Shooting result prediction model training method, storage medium and electronic device

Abstract

The application discloses a shooting result prediction model training method, a storage medium and an electronic device. The method comprises the following steps: acquiring a training sample set corresponding to a virtual shooting prop set, wherein each training sample in the training sample set comprises a shooting attribute feature set of one virtual shooting prop in the virtual shooting prop set and an actual shooting result obtained by using one virtual shooting prop, and the shooting attribute feature set comprises shooting attribute features in multiple dimensions; generating a group of classification tree models according to the training sample set; and selecting a target classification tree model in the group of classification tree models according to a target value set on a leaf node in the classification tree model and the actual shooting result in a training sample subset. The application can also be applied in an artificial intelligence scene and can also be related to machine learning and other technologies. The application solves the technical problem that a high-efficiency actual shooting result prediction means is lacked in related technologies.

Description

Technical Field

[0001] This invention relates to the field of computers, and more specifically, to a training method, storage medium, and electronic device for a shooting result prediction model. Background Technology

[0002] In recent years, shooting games have developed rapidly, and whether the strength of virtual shooting props will affect the balance of shooting games is a matter of great concern to both players and game staff.

[0003] An important indicator for determining whether virtual shooting items will affect game balance is their actual shooting results, such as the kill-to-death ratio (KD). However, since not all virtual shooting items are yet available for players to use in the game, it is necessary to predict the actual shooting results generated by virtual shooting items that have not yet been released for use, in order to determine whether these unreleased virtual shooting items will affect game balance in practice.

[0004] In related technologies, the prediction of actual shooting results largely depends on human intervention. However, human testing is highly susceptible to individual factors (such as the test-taker's skill level and energy levels). Therefore, the efficiency of related technologies in predicting actual shooting results cannot be guaranteed. In other words, related technologies lack an efficient means of predicting actual shooting results.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This invention provides a training method, storage medium, and electronic device for a shooting result prediction model, to at least address the technical problem of the lack of an efficient means of predicting actual shooting results in related technologies.

[0007] According to one aspect of the present invention, a method for training a shooting result prediction model is provided, comprising: obtaining a training sample set corresponding to a set of virtual shooting props, wherein each training sample in the training sample set includes a shooting attribute feature set of a virtual shooting prop in the set of virtual shooting props and an actual shooting result obtained using the virtual shooting prop, the shooting attribute feature set including shooting attribute features of multiple dimensions; generating a set of classification tree models based on the training sample set, wherein the non-leaf nodes of the classification tree models are set with shooting attribute features from the shooting attribute feature set and corresponding classification conditions, and the leaf nodes of the classification tree models are set with target values ​​of the actual shooting results in a subset of training samples, the target values ​​including the training attributes feature set. The average of the actual shooting results in the sample subset, the average residual of the difference between the actual shooting results in the training sample subset and the average of the above results, the shooting attribute feature set of each virtual shooting prop in the training sample subset satisfies the classification condition corresponding to the parent node of the above leaf node, and the training sample set includes the above training sample subset; based on the target value set on the leaf node in the above classification tree model and the actual shooting results in the above training sample subset, a target classification tree model is selected from the above set of classification tree models, wherein the target classification tree model is used to generate the predicted shooting result of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes and the corresponding classification conditions, and the target value set on the leaf nodes in the target classification tree model.

[0008] According to one aspect of the present invention, a method for predicting shooting results is provided, comprising: acquiring a set of target shooting attribute features of a virtual shooting prop to be predicted; determining a target leaf node corresponding to the virtual shooting prop in the leaf nodes of the target classification tree model according to the shooting attribute features set on non-leaf nodes in the target classification tree model and the corresponding classification conditions, and the set of target shooting attribute features, wherein the set of target shooting attribute features satisfies the classification conditions set on the parent node of the target leaf node; acquiring a target value set on the target leaf node, wherein the target value includes the average value of the actual shooting results and the average residual of the difference between the actual shooting results and the average value of the results; root Based on the target values ​​set on the leaf nodes of the aforementioned targets, the predicted shooting result of the aforementioned virtual shooting props is determined; wherein, the aforementioned target classification tree model is a classification tree model selected from the aforementioned set of classification tree models based on the target values ​​set on the leaf nodes of a set of classification tree models and the actual shooting results in the corresponding training sample subsets. The aforementioned set of classification tree models is a classification tree model generated based on the training sample set. Each training sample in the aforementioned training sample set includes a shooting attribute feature set of a virtual shooting prop in the virtual shooting prop set and an actual shooting result obtained by using the aforementioned virtual shooting prop. The shooting attribute feature set includes shooting attribute features in multiple dimensions. The aforementioned training sample set includes the aforementioned training sample subset.

[0009] According to another aspect of the present invention, a training apparatus for a shooting result prediction model is also provided, comprising: a first acquisition unit, configured to acquire a training sample set corresponding to a set of virtual shooting props, wherein each training sample in the training sample set includes a shooting attribute feature set of a virtual shooting prop in the set of virtual shooting props and an actual shooting result obtained using the virtual shooting prop, the shooting attribute feature set including shooting attribute features of multiple dimensions; and a generation unit, configured to generate a set of classification tree models based on the training sample set, wherein non-leaf nodes in the classification tree model are provided with shooting attribute features in the set of shooting attribute features and corresponding classification conditions, and leaf nodes in the classification tree model are provided with target values ​​of the actual shooting results in a subset of training samples, the target values ​​including... The training sample set includes the average of the actual shooting results in the above training sample subset and the average residual of the difference between the actual shooting results in the above training sample subset and the average of the above results. The shooting attribute feature set of each virtual shooting prop in the above training sample subset satisfies the classification condition corresponding to the parent node of the above leaf node. The training sample set includes the above training sample subset. The selection unit is used to select a target classification tree model from the above set of classification tree models based on the target value set on the leaf node in the above classification tree model and the actual shooting results in the above training sample subset. The target classification tree model is used to generate the predicted shooting result of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes and the corresponding classification conditions, and the target value set on the leaf nodes in the target classification tree model.

[0010] As an optional scheme, the above-mentioned generation unit includes: a first selection module, used to select a first shooting attribute feature in the shooting attribute feature set according to the feature value distribution of each shooting attribute feature in the shooting attribute feature set in the training sample set; a first generation module, used to generate a first classification tree model group with the node corresponding to the first shooting attribute feature as the root node, wherein the leaf nodes of the first classification tree model group are set with the above-mentioned average result, and the group of classification tree models includes the above-mentioned first classification tree model group; a second selection module, used to select a second shooting attribute feature in the shooting attribute feature set according to the feature value distribution of shooting attribute features other than the first shooting attribute feature in the shooting attribute feature set in the training sample set; a second generation module, used to generate a second classification tree model group with the node corresponding to the second shooting attribute feature as the root node, wherein the leaf nodes of the second classification tree model group are set with the above-mentioned average residual, and the group of classification tree models includes the above-mentioned second classification tree model group.

[0011] As an optional solution, the selection unit includes: a third selection module, used to select a first target classification tree model in the first classification tree model group based on the average result set on the leaf nodes of the first classification tree model group and the actual shooting results in the training sample subset, wherein the first target classification tree model is used to generate a first predicted shooting result of the target virtual shooting prop based on the shooting attribute features and corresponding classification conditions set on the non-leaf nodes of the first target classification tree model and the average result set on the leaf nodes; a fourth selection module, used to select a second target classification tree model in the second classification tree model group based on the average residual set on the leaf nodes of the second classification tree model group and the difference between the actual shooting results and the average result in the training sample subset, wherein the second target classification tree model is used to generate a second predicted shooting result of the target virtual shooting prop based on the shooting attribute features and corresponding classification conditions set on the non-leaf nodes of the second target classification tree model and the average residual set on the leaf nodes; wherein the predicted shooting result of the target virtual shooting prop is the sum of the first predicted shooting result and the second predicted shooting result.

[0012] As an optional solution, the selection module includes: a first determining submodule, configured to determine the classification condition corresponding to each shooting attribute feature in the shooting attribute feature set based on the feature value distribution of each shooting attribute feature in the shooting attribute feature set in the training sample set; and a first selection submodule, configured to select the first shooting attribute feature in the shooting attribute feature set based on the classification condition corresponding to each shooting attribute feature, wherein the plurality of training sample subsets are training sample subsets obtained by training samples in the training sample set according to the classification condition corresponding to the first shooting attribute feature, and the feature value of the first shooting attribute feature in each of the plurality of training sample subsets corresponds to a feature value cluster.

[0013] As an optional solution, the first generation module includes: a first setting submodule, used to set the node corresponding to the first shooting attribute feature as the root node in the first classification tree model to be generated; a first classification submodule, used to classify the training samples in the training sample set according to the classification conditions corresponding to the first shooting attribute feature to obtain multiple training sample subsets; a second setting submodule, used to set the node corresponding to each training sample subset in the multiple training sample subsets as a leaf node in the first classification tree model, wherein the leaf node in the first classification tree model is a child node of the root node in the first classification tree model, and the target value set on the leaf node in the first classification tree model is the average value of the actual shooting results in the corresponding training sample subset; and a first determining submodule, used to determine whether to generate a second classification tree model based on the error between the average value set on the leaf node in the first classification tree model and the actual shooting results in the corresponding training sample subset.

[0014] As an optional approach, the above-mentioned method of generating a second classification tree model group using the node corresponding to the second shooting attribute feature as the root node includes: a third setting submodule, used to set the node corresponding to the second shooting attribute feature as the root node in the third classification tree model to be generated; a second classification submodule, used to classify the training samples in the training sample set according to the classification conditions corresponding to the second shooting attribute feature to obtain multiple training sample subsets; a fourth setting submodule, used to set the node corresponding to each training sample subset in the multiple training sample subsets as a leaf node in the third classification tree model, wherein the leaf node in the third classification tree model is a child node of the root node in the third classification tree model, and the target value set on the leaf node in the third classification tree model is the average of the residual values ​​between the actual shooting results in the corresponding training sample subset and the average value of the actual shooting results in the training sample subset; and a second determining submodule, used to determine whether to generate a fourth classification tree model based on the error between the target value set on the leaf node in the third classification tree model and the residual value.

[0015] As an optional solution, the first determining submodule includes: a determining subunit, used to determine the generation of the second classification tree model when the error between the average value set on the leaf node of the first classification tree model and the actual shooting result in the corresponding training sample subset is greater than a third preset threshold, and to generate the second classification tree model based on the first classification tree model.

[0016] As an optional solution, the aforementioned determining subunit includes: a sub-determining module, used to determine a first leaf node in the aforementioned first classification tree model, wherein the error between the average value set on the aforementioned first leaf node and the actual shooting result in the training sample subset corresponding to the aforementioned first leaf node is greater than the aforementioned third preset threshold; a sub-selecting module, used to select a second shooting attribute feature in the aforementioned shooting attribute feature set according to the feature value distribution of shooting attribute features other than the aforementioned first shooting attribute feature in the aforementioned shooting attribute feature set in the aforementioned training sample set, wherein multiple training sample subsets are training sample subsets obtained by training samples in the aforementioned training sample set according to the classification conditions corresponding to the aforementioned second shooting attribute feature, and the feature value of the aforementioned second shooting attribute feature in each of the multiple training sample subsets corresponds to a feature value cluster; and a sub-setting module, used to replace the aforementioned first leaf node in the aforementioned first classification tree model with the child node corresponding to the aforementioned second shooting attribute feature, and set the node corresponding to each of the multiple training sample subsets as the leaf node of the child node corresponding to the aforementioned second shooting attribute feature, thereby obtaining the aforementioned second classification tree model.

[0017] As an optional scheme, the first generation module includes: an execution submodule, used to perform the following steps when the node depth in the i-th classification tree model is less than a fourth preset threshold, where i is a natural number; a third determination submodule, used to determine the generation of the (i+1)-th classification tree model when the error between the average value set on the leaf node in the i-th classification tree model and the actual shooting result in the corresponding training sample subset is greater than the third preset threshold; a third determination submodule, used to determine the first leaf node in the i-th classification tree model, where the error between the average value set on the first leaf node and the actual shooting result in the training sample subset corresponding to the first leaf node is greater than the third preset threshold; and a second selection submodule, used to select features from the shooting attribute feature set in the training sample set, excluding the above-mentioned... The feature value distribution of the shooting attribute features other than the first shooting attribute feature to the i-th shooting ray feature is selected from the above shooting attribute feature set. Among them, multiple training sample subsets are training sample subsets obtained by training samples in the above training sample set according to the classification conditions corresponding to the above i+1 shooting attribute feature. The feature value of the above i+1 shooting attribute feature in each of the above multiple training sample subsets corresponds to a feature value cluster. The replacement submodule is used to replace the above first leaf node in the above i-th classification tree model with the above i+1 child node corresponding to the above i+1 shooting attribute feature, and set the node corresponding to each of the above multiple training sample subsets as the leaf node of the above i+1 child node, so as to obtain the above i+1 classification tree model.

[0018] As an optional solution, it includes: an acquisition submodule, used to, after obtaining the above-mentioned (i+1)th classification tree model, acquire the error between the average value set on the leaf node of the above-mentioned (i+1)th child node and the actual shooting result in the training sample subset corresponding to the leaf node of the above-mentioned (i+1)th child node; and an adjustment submodule, used to, after obtaining the above-mentioned (i+1)th classification tree model, adjust the classification conditions corresponding to the above-mentioned (i+1)th shooting attribute feature set on the above-mentioned (i+1)th child node when there is a leaf node in the leaf node of the above-mentioned (i+1)th child node whose error is greater than the above-mentioned third preset threshold.

[0019] As an optional solution, the above selection module includes: a third selection submodule, used to select N shooting attribute features from the above shooting attribute feature set, where N is a natural number greater than 1; the above first generation module includes: a generation submodule, used to generate a set of classification tree models corresponding to each shooting attribute feature in parallel, with the node corresponding to one of the N shooting attribute features as the root node, resulting in a total of M sets of classification tree models, where the above set of classification tree models includes the above M sets of classification tree models, where M is a natural number greater than or equal to N and greater than 1.

[0020] As an optional approach, the above-mentioned generation submodule includes: a first generation subunit, used to generate in parallel a set of classification tree models corresponding to each of the above N shooting attribute features, with the node corresponding to one of the shooting attribute features as the root node, to obtain a total of M sets of classification tree models; or, a second generation subunit, used to generate in parallel a set of classification tree models corresponding to each of the above N shooting attribute features, with the node corresponding to one of the shooting attribute features as the root node and the nodes corresponding to one or more of the shooting attribute features as the child nodes of the root node, to obtain a total of M sets of classification tree models.

[0021] As an optional scheme, the above-mentioned generation unit includes: a fifth selection module, used to select the current shooting attribute feature from the above-mentioned shooting attribute feature set; a third generation module, used to generate a current classification tree model with the node corresponding to the current shooting attribute feature as the root node and the current classification condition corresponding to the current shooting attribute feature as the classification condition corresponding to the root node, wherein the leaf nodes in the above-mentioned current classification tree model group are set with current values, wherein the current values ​​include the average value of the results or the average value of the residuals; a sixth selection module, used to select the next shooting attribute feature from the above-mentioned shooting attribute feature set as the current shooting attribute feature when the error corresponding to the current value has not reached the convergence condition, or to use the next classification condition corresponding to the current shooting attribute feature as the current classification condition, or to select the next shooting attribute feature from the above-mentioned shooting attribute feature set as the current shooting attribute feature and use the classification condition corresponding to the next shooting attribute feature as the current classification condition; and a second determination module, used to determine that the above-mentioned group of classification tree models includes the above-mentioned current classification tree model when the error corresponding to the current value reaches the above-mentioned convergence condition.

[0022] According to another aspect of the present invention, a training apparatus for a shooting result prediction model is also provided, comprising: a second acquisition unit, configured to acquire a set of target shooting attribute features of a virtual shooting prop to be predicted; a first determination unit, configured to determine a target leaf node corresponding to the virtual shooting prop in the leaf nodes of the target classification tree model according to the shooting attribute features set on the non-leaf nodes of the target classification tree model and the corresponding classification conditions, and the set of target shooting attribute features, wherein the set of target shooting attribute features satisfies the classification conditions set on the parent node of the target leaf node; and a third acquisition unit, configured to acquire a target value set on the target leaf node, wherein the target value includes the average value of the actual shooting result and the average value of the actual shooting result and the average value of the result. The residual average of the difference; the second determining unit, used to determine the predicted shooting result of the virtual shooting prop of the target based on the target value set on the leaf node of the target; wherein, the target classification tree model is a classification tree model selected from the set of classification tree models based on the target value set on the leaf node of the set of classification tree models and the actual shooting result in the corresponding training sample subset, the set of classification tree models is a classification tree model generated based on the training sample set, each training sample in the training sample set includes a shooting attribute feature set of a virtual shooting prop in the virtual shooting prop set and an actual shooting result obtained by using the virtual shooting prop, the shooting attribute feature set includes shooting attribute features of multiple dimensions, and the training sample set includes the training sample subset.

[0023] As an optional solution, the third acquisition unit includes: a second acquisition module, used to acquire the average result and the average residual set on the leaf nodes of the target in the P classification tree models when the target classification tree model includes P classification tree models, wherein P is a natural number greater than 1; the second determination unit includes: a first determination module, used to determine the predicted shooting result of the target virtual shooting prop as equal to the sum of the average result and the average residual.

[0024] As an optional solution, the first determining unit includes: a search module, used to search for nodes in the target classification tree model that satisfy the classification conditions of the target shooting attribute feature set, starting from the root node in the target classification tree model, until the target leaf node is found.

[0025] According to another aspect of the present invention, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer program, wherein the computer program is configured to execute the training method of the above-described shooting result prediction model at runtime.

[0026] According to another aspect of the present invention, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the training method of the shooting result prediction model described above through the computer program.

[0027] In this embodiment of the invention, the shooting results of virtual shooting props are predicted using (classification tree) model technology to improve the automation of shooting result prediction. The highly automated prediction method can naturally improve the prediction efficiency of shooting results. Moreover, compared with manual testing in related technologies, model technology does not need to consider the influence of human factors and can stably and efficiently complete the shooting result prediction process.

[0028] Furthermore, the fit between model technology and the shooting result prediction scenario affects the accuracy of the prediction. In other words, simply applying model technology to the shooting result prediction scenario with low fit usually results in the model's output not meeting the accuracy requirements of the shooting result prediction scenario. Therefore, even if model technology has the advantage of higher automation compared to manual testing, if the output does not meet the accuracy requirements of the shooting result prediction scenario, it is not an efficient means of predicting shooting results. Therefore, in this embodiment of the invention, by setting shooting attribute features and corresponding classification conditions from the above-mentioned shooting attribute feature set on the non-leaf nodes of the classification tree model in the model technology, shooting attribute features and corresponding classification conditions that are more suitable for the shooting result prediction scenario are combined into the classification tree model. This makes the trained set of classification tree models a model with high fit with the shooting result prediction scenario. Then, the target classification tree model is selected from the set of classification tree models for actual shooting result prediction, ensuring the accuracy of the shooting result prediction output by the model.

[0029] In summary, this invention not only provides an efficient means of predicting actual shooting results through model technology, but also ensures the accuracy of actual shooting result prediction by combining a classification tree model with a shooting result prediction scenario with high consistency. Thus, it achieves the technical objective of improving the prediction efficiency of actual shooting results while ensuring the prediction accuracy of actual shooting results, thereby realizing the technical effect of providing an efficient and accurate means of predicting actual shooting results and solving the technical problem of the lack of an efficient means of predicting actual shooting results in related technologies. Attached Figure Description

[0030] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0031] Figure 1 This is a schematic diagram of the application environment of an optional shooting result prediction model training method according to an embodiment of the present invention;

[0032] Figure 2 This is a schematic diagram of the training method for an optional shooting result prediction model according to an embodiment of the present invention;

[0033] Figure 3 This is a schematic diagram of an optional shooting result prediction model training method according to an embodiment of the present invention;

[0034] Figure 4This is a schematic diagram of another optional training method for a shooting result prediction model according to an embodiment of the present invention;

[0035] Figure 5 This is a schematic diagram of another optional training method for a shooting result prediction model according to an embodiment of the present invention;

[0036] Figure 6 This is a schematic diagram of another optional training method for a shooting result prediction model according to an embodiment of the present invention;

[0037] Figure 7 This is a schematic diagram of another optional training method for a shooting result prediction model according to an embodiment of the present invention;

[0038] Figure 8 This is a schematic diagram of another optional training method for a shooting result prediction model according to an embodiment of the present invention;

[0039] Figure 9 This is a schematic diagram of another optional training method for a shooting result prediction model according to an embodiment of the present invention;

[0040] Figure 10 This is a schematic diagram of another optional training method for a shooting result prediction model according to an embodiment of the present invention;

[0041] Figure 11 This is a schematic diagram of another optional training method for a shooting result prediction model according to an embodiment of the present invention;

[0042] Figure 12 This is a schematic diagram of another optional training method for a shooting result prediction model according to an embodiment of the present invention;

[0043] Figure 13 This is a schematic diagram of the flow of an optional method for predicting shooting results according to an embodiment of the present invention;

[0044] Figure 14 This is a schematic diagram of a training device for an optional shooting result prediction model according to an embodiment of the present invention;

[0045] Figure 15 This is a schematic diagram of an optional shooting result prediction device according to an embodiment of the present invention;

[0046] Figure 16 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of the present invention. Detailed Implementation

[0047] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0048] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0049] First, to facilitate understanding of the embodiments of the present invention, some terms or nouns involved in the present invention will be explained below:

[0050] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0051] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0052] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0053] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, and smart customer service. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0054] The solutions provided in this application involve technologies such as machine learning in artificial intelligence, and are specifically illustrated through the following embodiments:

[0055] According to one aspect of the present invention, a method for training a shooting result prediction model is provided. Optionally, as an alternative implementation, the above-described method for training a shooting result prediction model can be applied, but is not limited to, to applications such as... Figure 1 The environment shown may include, but is not limited to, user equipment 102, network 110, and server 112. The user equipment 102 may include, but is not limited to, a display 108, a processor 106, and a memory 104.

[0056] The specific process can be summarized in the following steps:

[0057] Step S102: User equipment 102 obtains the prediction command triggered on the shooting result prediction interface 1022;

[0058] In steps S104-S106, user equipment 102 sends the prediction instruction to server 112 via network 110;

[0059] In step S108, server 112 searches for relevant historical data indicated by prediction command in database 114, and processes the relevant historical data through processing engine 116 to generate prediction shooting results.

[0060] In steps S110-S112, server 112 sends the predicted shooting results to user equipment 102 via network 110. Processor 106 in user equipment 102 displays the predicted shooting results on display 108 and stores the predicted shooting results in memory 104.

[0061] remove Figure 1 Beyond the illustrated example, the above steps can be performed independently by user equipment 102, i.e., user equipment 102 performs steps such as generating predicted shooting results, thereby reducing the processing load on the server. User equipment 102 includes, but is not limited to, handheld devices (such as mobile phones), laptops, desktop computers, in-vehicle devices, etc., and this invention does not limit the specific implementation of user equipment 102.

[0062] Alternatively, as an alternative implementation method, such as Figure 2 As shown, the training methods for the shooting result prediction model include:

[0063] S202, obtain the training sample set corresponding to the virtual shooting prop set, wherein each training sample in the training sample set includes the shooting attribute feature set of a virtual shooting prop in the virtual shooting prop set and the actual shooting result obtained by using a virtual shooting prop. The shooting attribute feature set includes shooting attribute features of multiple dimensions.

[0064] S204. Based on the training sample set, generate a set of classification tree models. The non-leaf nodes of the classification tree model are set with shooting attribute features from the shooting attribute feature set and corresponding classification conditions. The leaf nodes of the classification tree model are set with the target values ​​of the actual shooting results in the training sample subset. The target values ​​include the average value of the actual shooting results in the training sample subset and the average residual value of the difference between the actual shooting results and the average value in the training sample subset. The shooting attribute feature set of each virtual shooting prop in the training sample subset satisfies the classification conditions corresponding to the parent node of the leaf node. The training sample set includes the training sample subset.

[0065] S206. Based on the target values ​​set on the leaf nodes of the classification tree model and the actual shooting results in the training sample subset, select a target classification tree model from a set of classification tree models. The target classification tree model is used to generate the predicted shooting results of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes of the target classification tree model and the corresponding classification conditions, as well as the target values ​​set on the leaf nodes.

[0066] Optionally, in this embodiment, the training method of the shooting result prediction model described above can be applied, but is not limited to, in scenarios where the battle balance information of virtual shooting props is tested (used to measure whether the virtual shooting props disrupt the game balance of the virtual game in which the virtual shooting props are located due to excessively high / low prop strength). For example, a training sample set composed of multiple virtual shooting props as samples is first obtained, and each of the multiple virtual shooting props as samples carries historical usage data or is directly marked with one or more actual shooting results; then, the training sample set is used to train a classification tree model, and a set of classification tree models is obtained through training; furthermore, to ensure the accuracy of the model output, a target classification tree model is selected from the set of classification tree models as the actual prediction model;

[0067] Based on this, the item data (such as item type, item model, item accessories, etc.) of the virtual shooting item to be tested is input into the target classification tree model, and the output of the target classification tree model is used as the predicted shooting result of the virtual shooting item to be tested. Then, based on the obtained predicted shooting result, the balance of the virtual shooting item to be predicted is judged (the judgment result can be understood as balance information, but is not limited to). Based on the judgment result, the item data of the virtual shooting item to be predicted is adjusted (or not adjusted) to ensure that the virtual shooting item to be predicted will not disrupt the game balance of the virtual game it belongs to.

[0068] To further illustrate, alternatively, for example... Figure 3 As shown, assuming the virtual shooting prop to be predicted includes the following prop data: weapon type is rifle, model 25, accessories include a rear grip of model 1, a skill (GunPerk) of model 3, a scope of model 3, a grip of model 2, a muzzle of model 2, and a stock of model 1, then the above prop data is entered into the shooting result prediction interface 302. The shooting result prediction interface 302 also has virtual buttons "Get" and "Predict". The virtual button "Get" is used to trigger the acquisition of existing predicted shooting results (if the current virtual shooting prop to be predicted is being predicted for the first time, or there are no existing predicted shooting results, the virtual button "Get" cannot be triggered, or after triggering, a prompt message indicating that there are no existing predicted shooting results will be displayed). The virtual button "Predict" is used to trigger the prediction command to perform a prediction operation on the current virtual shooting prop to be predicted, such as... Figure 3 As shown in (a), selecting the virtual button "Predict" has triggered the prediction command;

[0069] Furthermore, such as Figure 3As shown in (b), in response to the prediction command, the predicted shooting result of the target virtual shooting prop is displayed in the prediction shooting result interface 304 (e.g., "0.9219936"). In addition, the prediction shooting result interface 304 is also equipped with a virtual button "Approve" and a virtual button "Cancel". The virtual button "Approve" is used to trigger the storage of the currently predicted shooting result in the existing prediction shooting results, and the virtual button "Cancel" is used to trigger the deletion of the currently predicted shooting result.

[0070] Optionally, in virtual shooting games where virtual shooting props are located, there are often many different types of virtual shooting prop main bodies, and each virtual shooting prop main body can be used as a virtual shooting prop to test or predict shooting results. In addition, virtual shooting games often have many different types of virtual shooting prop accessories, and virtual shooting prop main bodies and virtual shooting prop accessories can be flexibly combined to form new virtual shooting props. This means that testing or predicting shooting results requires not only testing all types of virtual shooting prop main bodies, but also testing new virtual shooting props obtained through any combination method, which dramatically increases the number of objects to be predicted or tested, and the pressure of prediction or testing also increases.

[0071] Optionally, in this embodiment, whether it is a virtual shooting prop body that is used alone as a virtual shooting prop, or a virtual shooting prop body and virtual shooting prop accessories that are used in combination as virtual shooting props, they can be regarded as prop data with the same or different shooting attribute characteristics, without distinguishing the type or combination method. Therefore, even if there are a large number of objects to be predicted or tested, there is no prediction or testing pressure in this embodiment.

[0072] Optionally, in this embodiment, the shooting attribute features may be, but are not limited to, attribute features in the shooting game scene that are highly correlated with the actual shooting results, such as the damage multiplier of a virtual bullet hitting the head of an enemy virtual object, the damage multiplier of a virtual bullet hitting the torso of an enemy virtual object, the firing interval (such as the time interval between two bullets firing in burst mode), the vertical recoil of the shot (base value, modification factor, maximum value), the horizontal recoil of the shot (base value, modification factor, maximum value), the time to put away the virtual shooting prop (the time required to raise the gun and put it down), and the time to raise the virtual shooting prop, etc.

[0073] Optionally, in this embodiment, the actual shooting result can be used, but is not limited to, to represent the combat result fed back during one or more rounds of shooting games using virtual shooting props, such as kill result, kill result, assist result, etc. Furthermore, predicting the shooting result of a virtual shooting prop can be understood, but is not limited to, predicting the combat result that the virtual shooting prop should feed back in each future round of shooting games.

[0074] Optionally, in this embodiment, the classification tree model may include, but is not limited to, a root node (non-leaf node), one or more leaf nodes, and one or more child nodes (non-leaf nodes). The root node can be understood as an input node, that is, classifying the shooting attribute features input to the classification tree model and outputting the classification result to the next node (which can be a child node or a leaf node). If the next node is a corresponding leaf node, a preset value of the corresponding leaf node is output, and the output result of the current classification tree model is determined based on this preset data. Optionally, in the classification tree model, each non-leaf node may have, but is not limited to, a preset classification condition, and this corresponding classification condition is used to classify the input features for output to the next node.

[0075] Optionally, in this embodiment, a set of classification tree models includes multiple classification tree models, and the target classification tree model selected from the set of classification tree models may include, but is not limited to, multiple classification tree models. In other words, the selection criteria in a set of classification tree models may include, but is not limited to, selecting all classification tree models that meet the preset conditions as the target classification tree model; furthermore, the target classification tree model selected from a set of classification tree models may include, but is not limited to, a single classification tree model. In other words, the selection criteria in a set of classification tree models may include, but is not limited to, selecting the classification tree model that best meets the preset conditions as the target classification tree model.

[0076] It should be noted that for ordinary prediction scenarios, model technology is undoubtedly the best way to ensure the accuracy and efficiency of prediction. However, the prediction scenarios of actual shooting results are different from ordinary prediction scenarios. The prediction scenarios of actual shooting results have high scenario complexity (there are many and complex factors affecting the actual shooting results, such as the rate of fire will change at different ranges, and the damage caused by the shooting will also change depending on the hit site). Therefore, how to apply model technology to the prediction scenarios of actual shooting results is still a difficult problem that related technologies have not yet overcome.

[0077] In this embodiment, to overcome this challenge, a classification tree model in model technology is proposed to predict actual shooting results. Furthermore, by setting shooting attribute features and corresponding classification conditions from the shooting attribute feature set on the non-leaf nodes of the classification tree model, shooting attribute features and corresponding classification conditions that closely match the shooting result prediction scenario are combined into the classification tree model, resulting in a set of classification tree models with a high degree of fit with the shooting result prediction scenario. The training process of the classification tree model is further improved by setting the average value of actual shooting results from the training sample subset on the leaf nodes. Finally, a target classification tree model is selected from the set of classification tree models for actual shooting result prediction, ensuring the accuracy of the shooting result prediction output by the model.

[0078] Further examples, such as Figure 4 As shown, firstly, the root node 1-1 in the classification tree model 11 is constructed using all or part of the shooting attribute features in the shooting attribute feature set of the training sample set 404, and the classification condition 402 set on the root node 1-1 is constructed. From this, two leaf nodes with the root node 1-1 as the parent node can be obtained. Then, each training sample in the training sample set is input into the root node 1-1 in sequence, and the classification condition 402 is used to classify each input training sample. The classification result is to be classified into leaf node 1-2 or leaf node 1-3.

[0079] After classifying each training sample using classification condition 402, the training sample subsets containing leaf nodes 1-2 and 1-3 are statistically analyzed, and the average value (average result) of the actual shooting results recorded in the training sample set 404 for each training sample subset is calculated. This average value is then used as the weight value of the corresponding leaf node.

[0080] Furthermore, optionally, after obtaining the average value of the actual shooting results (sum) corresponding to each training sample subset, the average value can be compared with each actual shooting result recorded in the training sample set 404 for each training sample subset to obtain the corresponding error. If the error does not meet the preset conditions, the weight values ​​of classification condition 402 and / or leaf nodes can be adjusted. That is, the determined weight values ​​of classification condition 402 and / or leaf nodes can be, but are not limited to, adjustable; and the weight values ​​of classification condition 402 and / or leaf nodes can be set to non-adjustable after the classification tree model 11 has been trained.

[0081] Furthermore, after the classification tree model 11 has been trained, it is possible, but not limited to, to continue generating the next classification tree model, namely classification tree model 12. The generation of classification tree model 12 is based on the root node 1-1 and leaf node 1-3 of classification tree model 11, and then uses all or part of the shooting attribute features in the shooting attribute feature set to construct child node 1-2 and classification condition 404, thereby obtaining two leaf nodes with child node 1-2 as the parent node, namely leaf node 1-4 and leaf node 1-5;

[0082] Optionally, in the process of generating classification tree model 12 based on classification tree model 11, the child nodes corresponding to the node positions of leaf nodes 1-2 and / or leaf nodes 1-3 can be selected as the parent nodes of the leaf nodes of classification tree model 12, and the selection criteria can be, but are not limited to, the errors corresponding to leaf nodes 1-2 and leaf nodes 1-3. For example, the child nodes corresponding to the node positions of leaf nodes with larger errors can be selected as the parent nodes of the leaf nodes of classification tree model 12; or, the child nodes corresponding to the node positions of leaf nodes with errors exceeding the error threshold can be selected as the parent nodes of the leaf nodes of classification tree model 12.

[0083] Furthermore, based on the same logic, the nth classification tree model, i.e., classification tree model 1n, is generated. Among the n classification tree models, the one with the smallest error is selected as the first target tree model, which is used to generate the initial predicted shooting result of the target virtual shooting prop. Although the error of the first target tree model is relatively the smallest, residuals inevitably still exist. Therefore, it is necessary to use the average result corresponding to each leaf node in the first target tree model to calculate the difference with the actual shooting result corresponding to each leaf node to obtain multiple differences. Then, these multiple differences are summed and averaged to obtain the average residual.

[0084] Optionally, in this embodiment, based on Figure 4 The scenario shown continues, for example... Figure 5 As shown, assuming that the classification tree model 2 has been determined as the first target tree model 502, the training sample subset corresponding to each leaf node is obtained, and the actual shooting result of the training sample subset corresponding to each leaf node is modified to the difference between the average result and the actual shooting result. Alternatively, each training sample recorded in the original training sample set 404 records the actual shooting result as the sample label, but in this embodiment, after determining the first target classification tree model 502, the difference between the average result and the actual shooting result is used as the sample label for each training sample, thereby obtaining a new training sample set 506.

[0085] Furthermore, the root node 2-1 in the classification tree model 21 is constructed using all or part of the shooting attribute features in the shooting attribute feature set of the training sample set 506, along with the classification condition 502 set on the root node 2-1. Thus, two leaf nodes with the root node 2-1 as the parent node can be obtained. Then, each training sample in the training sample set 502 is sequentially input into the root node 2-1, and the classification condition 502 is used to classify each input training sample. The classification result is assigned to the leaf node 2-2 or the leaf node 2-3.

[0086] After classifying each training sample using classification condition 502, the training sample subsets containing leaf nodes 2-2 and 2-3 are statistically analyzed, and the average difference between the average result recorded in training sample set 502 and the actual shooting result (residual average) of each training sample subset is calculated. Then, the residual average is used as the weight value of the corresponding leaf node.

[0087] Furthermore, optionally, after obtaining the residual average value corresponding to each training sample subset, the residual average value can be compared with the difference between the average value of each result recorded in the training sample set 502 and the actual shooting result for each training sample subset to obtain the corresponding error. If the error does not meet the preset conditions, the weight values ​​of classification condition 502 and / or leaf nodes can be adjusted. That is, the determined weight values ​​of classification condition 502 and / or leaf nodes can be, but are not limited to, adjustable; and the weight values ​​of classification condition 502 and / or leaf nodes can be, but are not limited to, set to non-adjustable after the classification tree model 21 has been trained.

[0088] Furthermore, after the classification tree model 21 has been trained, it is possible, but not limited to, to continue generating the next classification tree model, namely classification tree model 22. The generation of classification tree model 22 is based on the root node 2-1 and leaf node 23 of classification tree model 21, and then uses all or part of the shooting attribute features in the shooting attribute feature set to construct child node 2-2 and classification condition 504, and thus obtains two leaf nodes with child node 2-2 as the parent node, namely leaf node 2-4 and leaf node 2-5;

[0089] Optionally, in the process of generating classification tree model 22 based on classification tree model 21, the child nodes corresponding to the node positions of leaf node 2-2 and / or leaf node 2-3 can be selected as the parent nodes of the leaf nodes of classification tree model 22, and the selection criteria can be, but are not limited to, the errors corresponding to leaf node 2-2 and leaf node 2-3. For example, the child nodes corresponding to the node positions of leaf nodes with larger errors can be selected as the parent nodes of the leaf nodes of classification tree model 22; or, the child nodes corresponding to the node positions of leaf nodes with errors exceeding the error threshold can be selected as the parent nodes of the leaf nodes of classification tree model 22.

[0090] Furthermore, based on Figure 4 The same logic in the illustrated embodiment generates the qth classification tree model, namely classification tree model 2q, and selects the classification tree model with the smallest error among the q classification tree models as the second target tree model, which is used to fit the residual existing in the first target tree model 502.

[0091] Furthermore, assuming that classification tree model 2q is determined as the second target classification tree model 510, then classification tree model 12 (first target classification tree model 502) and classification tree model 2q (second target classification tree model 510) can be used together as the target classification tree model 510, but not limited to.

[0092] Optionally, in this embodiment, a set of classification tree models may include, but is not limited to, including... Figure 4 The n classification tree models trained using a training sample set of 404 can also include, but are not limited to, the following: Figure 5 The training sample set consists of 506 training trees, resulting in q classification tree models. Furthermore, the number of classification tree models in a set, n, q, or (n+q), can be determined by one or more constraints, including but not limited to. For example, the maximum number of iterations for each classification tree model can be pre-limited, for instance, set to n, meaning at most n trees will be generated during training, i.e., n iterations. Another example is pre-limiting the maximum tree depth, for instance, setting it to 10 layers, meaning the depth of each tree cannot exceed 10 layers during training. Yet another example is pre-setting convergence conditions, such that if any tree's output satisfies these convergence conditions during training, training ends, or in other words, the currently generated classification tree model is determined to be a set of classification tree models.

[0093] To further illustrate, the optional assumptions have already been... Figure 5 If a target classification tree model is selected from the set of classification tree models shown, then the process of generating the predicted shooting result of the target virtual shooting prop can be, but is not limited to, as follows: Figure 6As shown, firstly, the shooting item data 602 of the target virtual shooting item 604 is obtained, and then the shooting item data 602 of the target virtual shooting item 604 is input into the target classification tree model 606; the target classification tree model 606 is as follows. Figure 6 As shown, the tree structure includes three layers. The first layer includes a root node 1-1, which has a classification condition 606-1. The second layer includes a child node 2-1 and leaf nodes 1-6, where the child node 2-1 has a classification condition 606-2, and the weight of leaf node 1-6 is assumed to be 0.6. The third layer includes a leaf node 3-1 and a leaf node 3-2, with the weight of leaf node 3-1 assumed to be 0.3 and the weight of leaf node 3-2 assumed to be 0.2.

[0094] Furthermore, based on the matching degree between shooting prop data 602 and classification condition 606-1, the target virtual shooting prop 604 is classified at the root node 1-1. Assuming the target virtual shooting prop 604 is classified to leaf node 1-6, the output predicted shooting result 608 is the weight value of leaf node 1-6, which is 0.6. Conversely, if the target virtual shooting prop 604 is classified to child node 2-1, then based on the matching degree between shooting prop data 602 and classification condition 606-2, the target virtual shooting prop 604 is classified at child node 2-1. If the target virtual shooting prop 604 is classified to leaf node 3-1, the output predicted shooting result 608 is the weight value of leaf node 3-1, which is 0.3 (target value); or if the target virtual shooting prop 604 is classified to leaf node 3-2, the output predicted shooting result 608 is the weight value of leaf node 3-2, which is 0.2.

[0095] Furthermore, in this embodiment, it is also possible, but not limited to, the following: Figure 7 As shown, firstly, the shooting item data 702 of the target virtual shooting item 704 is obtained, and then the shooting item data 702 of the target virtual shooting item 704 is input into the target classification tree model (including the first target classification model 706 and the second target classification model 708); wherein, the first target classification model 706 is as follows Figure 7 As shown, the tree structure includes three layers. The first layer includes a root node 1-1, which has a classification condition 706-1. The second layer includes a child node 1-1 and a leaf node 1-2, where the child node 1-1 has a classification condition 706-2, and the weight of the leaf node 1-2 is assumed to be 0.7. The third layer includes a leaf node 1-3 and a leaf node 1-3, with the weights of the leaf nodes 1-3 and 1-3 respectively being 0.3 and 0.2.

[0096] Furthermore, the second target classification model 708, such asFigure 7 As shown, there are two tree structures. The first tree structure includes a root node 2-1, where the root node 2-1 is set with classification condition 708-1. The second tree structure includes leaf node 2-1 and leaf node 2-2, where it is assumed that the weight of leaf node 2-1 is -0.2 and the weight of leaf node 2-2 is -0.1.

[0097] Furthermore, based on the matching degree between the shooting prop data 702 and the classification condition 707-1, the target virtual shooting prop 704 is classified at root node 1-1 and root node 2-1. Assuming that the target virtual shooting prop 704 is classified to leaf node 1-2 and leaf node 2-1, the output predicted shooting result 708 is the sum of the weight value of leaf node 1-2 (0.7) and the weight value of leaf node 2-1 (-0.2), which is 0.5. Conversely, if the target virtual shooting prop 704 is classified to child node 1-1, then based on the matching degree between the shooting prop data 702 and the classification condition 707-2, the target virtual shooting prop 704 is classified at child node 2-1. If the target virtual shooting prop 704 is classified to leaf node 1-3 and leaf node 2-2, the output predicted shooting result 708 is the sum of the weight value of leaf node 1-3 (0.3) and the weight value of leaf node 2-2 (-0.1), which is 0.2.

[0098] The embodiments provided in this application obtain a training sample set corresponding to a set of virtual shooting props. Each training sample in the training sample set includes a shooting attribute feature set of a virtual shooting prop from the set and the actual shooting result obtained by using a virtual shooting prop. The shooting attribute feature set includes shooting attribute features of multiple dimensions. Based on the training sample set, a set of classification tree models is generated. The non-leaf nodes of the classification tree model are set with shooting attribute features from the shooting attribute feature set and corresponding classification conditions. The leaf nodes of the classification tree model are set with target values ​​of the actual shooting results in the training sample subset. The target values ​​include the average result of the actual shooting results in the training sample subset and the average residual of the difference between the actual shooting results and the average result in the training sample subset. The shooting attribute feature set of each virtual shooting prop in the training sample subset satisfies the leaf node... The classification conditions corresponding to the parent node are defined, and the training sample set includes a subset of training samples. Based on the target values ​​set on the leaf nodes of the classification tree model and the actual shooting results in the subset of training samples, a target classification tree model is selected from a set of classification tree models. The target classification tree model is used to generate the predicted shooting results of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes of the target classification tree model, the corresponding classification conditions, and the target values ​​set on the leaf nodes. This model technology provides an efficient means of predicting actual shooting results. Furthermore, by combining the classification tree model with the shooting result prediction scenario with a high degree of fit, the accuracy of the actual shooting result prediction is also guaranteed. Thus, the technical objective of improving the prediction efficiency of actual shooting results while ensuring the prediction accuracy of actual shooting results is achieved, thereby realizing the technical effect of providing an efficient and accurate means of predicting actual shooting results.

[0099] As an optional approach, a set of classification tree models is generated based on the training sample set, including:

[0100] S1, based on the feature value distribution of each shooting attribute feature in the shooting attribute feature set in the training sample set, select the first shooting attribute feature in the shooting attribute feature set;

[0101] S2, using the node corresponding to the first shooting attribute feature as the root node, generate the first classification tree model group, wherein the leaf nodes in the first classification tree model group are set with the average result, and a group of classification tree models includes the first classification tree model group;

[0102] S3, based on the feature value distribution of the shooting attribute features other than the first shooting attribute feature in the shooting attribute feature set in the training sample set, select the second shooting attribute feature from the shooting attribute feature set;

[0103] S4. Using the node corresponding to the second shooting attribute feature as the root node, generate a second classification tree model group. The leaf nodes in the second classification tree model group are set with residual average values. A set of classification tree models includes the second classification tree model group.

[0104] Optionally, in this embodiment, the feature distribution may, but is not limited to, represent the distribution of the same type of features. For example, assuming the training sample set includes 100 samples, and each sample includes 3 types of features (such as type A, type B, and type C), then the feature distribution may, but is not limited to, represent the distribution of a single type of feature in all samples (100). Alternatively, in this scenario, the distribution of type A features, type B features, and type C features in all samples (100) needs to be obtained separately, and the first shooting attribute feature is selected based on the distribution. For example, based on the distribution, type A features are selected as the first shooting attribute feature from type A features, type B features, and type C features.

[0105] Optionally, in this embodiment, the first classification tree model group can be understood, but is not limited to, as Figure 5 The diagram shows n classification tree models trained using a training sample set of 404; the second classification tree model group can be understood, but is not limited to, as... Figure 5 The image shows q classification tree models trained using a training sample set of 506.

[0106] As an optional approach, based on the target values ​​set on the leaf nodes of the classification tree model and the actual shooting results in the training sample subset, a target classification tree model is selected from a set of classification tree models, including:

[0107] S1, based on the average result set on the leaf nodes in the first classification tree model group and the actual shooting results in the training sample subset, select the first target classification tree model in the first classification tree model group. The first target classification tree model is used to generate the first predicted shooting result of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes in the first target classification tree model and the corresponding classification conditions, as well as the average result set on the leaf nodes.

[0108] S2, based on the residual average value set on the leaf nodes in the second classification tree model group and the difference between the actual shooting results and the average value of the results in the training sample subset, select the second target classification tree model in the second classification tree model group. The second target classification tree model is used to generate the second predicted shooting result of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes in the second target classification tree model and the corresponding classification conditions, as well as the residual average value set on the leaf nodes.

[0109] It should be noted that a set of classification tree models can be understood as, but is not limited to, two sets or at least two sets of classification tree models, and a target classification tree model can be understood as, but is not limited to, determining the target classification tree model corresponding to each set of classification tree models in the two sets or at least two sets of classification tree models, namely, the first target classification tree model and the second target classification tree model. The second target classification tree model does not impose a limit on the number of other target classification tree models besides the first target classification tree model.

[0110] The predicted firing result of the target virtual shooting prop is the sum of the first predicted firing result and the second predicted firing result.

[0111] Further examples, such as Figure 4 As shown, firstly, the root node 1-1 in the classification tree model 11 is constructed using all or part of the shooting attribute features in the shooting attribute feature set of the training sample set 404, and the classification condition 402 set on the root node 1-1 is constructed. From this, two leaf nodes with the root node 1-1 as the parent node can be obtained. Then, each training sample in the training sample set is input into the root node 1-1 in sequence, and the classification condition 402 is used to classify each input training sample. The classification result is to be classified into leaf node 1-2 or leaf node 1-3.

[0112] After classifying each training sample using classification condition 402, the training sample subsets containing leaf nodes 1-2 and 1-3 are statistically analyzed, and the average value (average result) of the actual shooting results recorded in the training sample set 404 for each training sample subset is calculated. This average value is then used as the weight value of the corresponding leaf node.

[0113] Furthermore, optionally, after obtaining the average value of the actual shooting results (sum) corresponding to each training sample subset, the average value can be compared with each actual shooting result recorded in the training sample set 404 for each training sample subset to obtain the corresponding error. If the error does not meet the preset conditions, the weight values ​​of classification condition 402 and / or leaf nodes can be adjusted. That is, the determined weight values ​​of classification condition 402 and / or leaf nodes can be, but are not limited to, adjustable; and the weight values ​​of classification condition 402 and / or leaf nodes can be set to non-adjustable after the classification tree model 11 has been trained.

[0114] Furthermore, after the classification tree model 11 has been trained, it is possible, but not limited to, to continue generating the next classification tree model, namely classification tree model 12. The generation of classification tree model 12 is based on the root node 1-1 and leaf node 1-3 of classification tree model 11, and then uses all or part of the shooting attribute features in the shooting attribute feature set to construct child node 1-2 and classification condition 404, thereby obtaining two leaf nodes with child node 1-2 as the parent node, namely leaf node 1-4 and leaf node 1-5;

[0115] Optionally, in the process of generating classification tree model 12 based on classification tree model 11, the child nodes corresponding to the node positions of leaf nodes 1-2 and / or leaf nodes 1-3 can be selected as the parent nodes of the leaf nodes of classification tree model 12, and the selection criteria can be, but are not limited to, the errors corresponding to leaf nodes 1-2 and leaf nodes 1-3. For example, the child nodes corresponding to the node positions of leaf nodes with larger errors can be selected as the parent nodes of the leaf nodes of classification tree model 12; or, the child nodes corresponding to the node positions of leaf nodes with errors exceeding the error threshold can be selected as the parent nodes of the leaf nodes of classification tree model 12.

[0116] Furthermore, based on the same logic, the nth classification tree model, i.e., classification tree model 1n, is generated. Among the n classification tree models, the one with the smallest error is selected as the first target tree model, which is used to generate the initial predicted shooting result of the target virtual shooting prop. Although the error of the first target tree model is relatively the smallest, residuals inevitably still exist. Therefore, it is necessary to use the average result corresponding to each leaf node in the first target tree model to calculate the difference with the actual shooting result corresponding to each leaf node to obtain multiple differences. Then, these multiple differences are summed and averaged to obtain the average residual.

[0117] Optionally, in this embodiment, based on Figure 4 The scenario shown continues, for example... Figure 5 As shown, assuming that the classification tree model 2 has been determined as the first target tree model 502, the training sample subset corresponding to each leaf node is obtained, and the actual shooting result of the training sample subset corresponding to each leaf node is modified to the difference between the average result and the actual shooting result. Alternatively, each training sample recorded in the original training sample set 404 records the actual shooting result as the sample label, but in this embodiment, after determining the first target classification tree model 502, the difference between the average result and the actual shooting result is used as the sample label for each training sample, thereby obtaining a new training sample set 506.

[0118] Furthermore, the root node 2-1 in the classification tree model 21 is constructed using all or part of the shooting attribute features in the shooting attribute feature set of the training sample set 506, along with the classification condition 502 set on the root node 2-1. Thus, two leaf nodes with the root node 2-1 as the parent node can be obtained. Then, each training sample in the training sample set 502 is sequentially input into the root node 2-1, and the classification condition 502 is used to classify each input training sample. The classification result is assigned to the leaf node 2-2 or the leaf node 2-3.

[0119] After classifying each training sample using classification condition 502, the training sample subsets containing leaf nodes 2-2 and 2-3 are statistically analyzed, and the average difference between the average result recorded in training sample set 502 and the actual shooting result (residual average) of each training sample subset is calculated. Then, the residual average is used as the weight value of the corresponding leaf node.

[0120] Furthermore, optionally, after obtaining the residual average value corresponding to each training sample subset, the residual average value can be compared with the difference between the average value of each result recorded in the training sample set 502 and the actual shooting result for each training sample subset to obtain the corresponding error. If the error does not meet the preset conditions, the weight values ​​of classification condition 502 and / or leaf nodes can be adjusted. That is, the determined weight values ​​of classification condition 502 and / or leaf nodes can be, but are not limited to, adjustable; and the weight values ​​of classification condition 502 and / or leaf nodes can be, but are not limited to, set to non-adjustable after the classification tree model 21 has been trained.

[0121] Furthermore, after the classification tree model 21 has been trained, it is possible, but not limited to, to continue generating the next classification tree model, namely classification tree model 22. The generation of classification tree model 22 is based on the root node 2-1 and leaf node 23 of classification tree model 21, and then uses all or part of the shooting attribute features in the shooting attribute feature set to construct child node 2-2 and classification condition 504, and thus obtains two leaf nodes with child node 2-2 as the parent node, namely leaf node 2-4 and leaf node 2-5;

[0122] Optionally, in the process of generating classification tree model 22 based on classification tree model 21, the child nodes corresponding to the node positions of leaf node 2-2 and / or leaf node 2-3 can be selected as the parent nodes of the leaf nodes of classification tree model 22, and the selection criteria can be, but are not limited to, the errors corresponding to leaf node 2-2 and leaf node 2-3. For example, the child nodes corresponding to the node positions of leaf nodes with larger errors can be selected as the parent nodes of the leaf nodes of classification tree model 22; or, the child nodes corresponding to the node positions of leaf nodes with errors exceeding the error threshold can be selected as the parent nodes of the leaf nodes of classification tree model 22.

[0123] Furthermore, based on Figure 4 The same logic in the illustrated embodiment generates the qth classification tree model, namely classification tree model 2q, and selects the classification tree model with the smallest error among the q classification tree models as the second target tree model, which is used to fit the residual existing in the first target tree model 502.

[0124] Furthermore, assuming that classification tree model 2q is determined as the second target classification tree model 510, then classification tree model 12 (first target classification tree model 502) and classification tree model 2q (second target classification tree model 510) can be used together as the target classification tree model 510, but not limited to.

[0125] As an optional approach, based on the feature value distribution of each shooting attribute feature in the shooting attribute feature set within the training sample set, a first shooting attribute feature is selected from the shooting attribute feature set, including:

[0126] S1, Based on the feature value distribution of each shooting attribute feature in the shooting attribute feature set in the training sample set, determine the classification conditions corresponding to each shooting attribute feature in the shooting attribute feature set;

[0127] S2, select the first shooting attribute feature from the shooting attribute feature set according to the classification conditions corresponding to each shooting attribute feature. The multiple training sample subsets are training sample subsets obtained by training samples in the training sample set according to the classification conditions corresponding to the first shooting attribute feature. The feature value of the first shooting attribute feature in each of the multiple training sample subsets corresponds to a feature value cluster.

[0128] Optionally, in this embodiment, the feature value distribution can not only be used as the basis for constructing the root node, but also, but is not limited to, as the basis for determining the classification conditions corresponding to the root node.

[0129] To further illustrate, consider an optional assumption that the training sample set includes 100 samples, each containing three types of features (e.g., A, B, and C). Taking A-class features as an example, we obtain one or more splitting points corresponding to the segmentation operation of the 100 samples under A-class features. For instance, if the 100 A-class features include 1, 2, 3…100, then when segmenting into 1 and the other 99 A-class features, the splitting point is greater than or equal to 1. Based on the same logic, we can summarize and obtain 99 splitting points. Further, we perform classification based on these 99 splitting points, and then calculate the corresponding splitting point index (e.g., an uncertainty index indicating whether the segmented feature clusters are uncertain, or a discriminant index measuring whether features are clearly distinguished). Then, based on this splitting point index, we select one splitting point from the above 99 splitting points as the optimal splitting point for A-class features. Furthermore, the calculation method for the splitting point index can, but is not limited to, refer to the feature value clusters corresponding to each splitting point.

[0130] Furthermore, based on the same logic, the optimal segmentation points for Class B features and Class C features are obtained, and the segmentation point indices of these three optimal segmentation points are compared to select the segmentation point corresponding to the optimal segmentation point index (assuming it is the optimal segmentation point for Class A features). Then, Class A features are used as the first shooting attribute features, and the classification conditions corresponding to the first shooting attribute features are determined based on the optimal segmentation point of Class A features (assuming the optimal segmentation point is greater than 1 or less than or equal to 1, then the classification condition is binary classification based on greater than 1 or less than or equal to 1).

[0131] It should be noted that, for ease of understanding, the above embodiments use binary classification as an example, but this does not mean that this application does not involve other classification methods besides binary classification, and there is no limitation on them.

[0132] To further illustrate, the classification criteria can also be multiple selection ranges, such as [0, 1], (1, 2], (2, 3]; and correspondingly, the number of leaf nodes or non-leaf nodes under the root node is no longer limited to 2, but is based on the number of selection ranges corresponding to the above classification criteria. For example, if the number of the above selection ranges is 3, then the number of leaf nodes or non-leaf nodes under the root node is also 3.

[0133] The embodiments provided in this application determine the classification conditions corresponding to each shooting attribute feature in the shooting attribute feature set based on the feature value distribution of each shooting attribute feature in the shooting attribute feature set in the training sample set. Based on the classification conditions corresponding to each shooting attribute feature, a first shooting attribute feature is selected in the shooting attribute feature set. The multiple training sample subsets are training sample subsets obtained by training samples in the training sample set according to the classification conditions corresponding to the first shooting attribute feature. The feature value of the first shooting attribute feature in each of the multiple training sample subsets corresponds to a feature value cluster. This achieves the purpose of using features with high discriminative power as the root node and the basis for determining the classification conditions, thereby improving the efficiency of model training.

[0134] As an optional approach, the first classification tree model group is generated using the node corresponding to the first shooting attribute feature as the root node, including:

[0135] S1, set the node corresponding to the first shooting attribute feature as the root node in the first classification tree model to be generated;

[0136] S2, classify the training samples in the training sample set according to the classification conditions corresponding to the first shooting attribute features to obtain multiple training sample subsets;

[0137] S3, set the node corresponding to each training sample subset in the multiple training sample subsets as a leaf node in the first classification tree model, wherein the leaf node in the first classification tree model is a child node of the root node in the first classification tree model, and the target value set on the leaf node in the first classification tree model is the average value of the actual shooting results in the corresponding training sample subset.

[0138] S4. Based on the error between the target value set on the leaf node of the first classification tree model and the actual shooting result in the corresponding training sample subset, determine whether to generate the second classification tree model.

[0139] It should be noted that the training process of a set of classification tree models can be understood as training in two concepts: vertical and horizontal. The vertical concept refers to the process of training a single classification tree model, including how to construct the root node, leaf nodes, or other non-leaf nodes, and how to set the corresponding classification conditions on each non-leaf node. The horizontal concept can be understood as the process of training multiple classification tree models, i.e., whether to train the next classification tree model after the training of one classification tree model is completed, and if so, how to train the next classification tree model.

[0140] To further illustrate, optional examples include... Figure 8As shown, each training sample in the training sample set 802 is sequentially input into the initially constructed classification tree model 804 (the first classification tree model), and classification is first performed at the root node 1-1 according to the classification condition 806; then, training sample subsets 808-1 classified to leaf node 2-1 and 808-2 classified to leaf node 2-2 are obtained; the target value corresponding to each training sample in training sample subset 808-1 is calculated, and then the average value is used to calculate the difference for each actual shooting result, finally obtaining the error 810-1 of training sample subset 808-1 at leaf node 2-1; similarly, the average value of the sum of the actual shooting results corresponding to each training sample in training sample subset 808-2 is calculated, and then the average value is used to calculate the difference for each actual shooting result, finally obtaining the error 810-2 of training sample subset 808-2 at leaf node 2-2;

[0141] Furthermore, based on errors 810-1 and 810-2, leaf nodes 2-1 and 2-2 are judged respectively. If error 810-1 exceeds the threshold, it indicates that the current weight value of leaf node 2-1 (i.e., the average of the sum of the actual shooting results corresponding to each training sample in the training sample subset 808-1) has a residual and needs to be fitted to meet the generation conditions of the next classification tree model. In the training process of the next classification tree model, leaf node 1-4 is used as the parent node of the next leaf node to complete the residual fitting. Similarly, if the error If error 810-2 exceeds the threshold, it means that the current weight value of leaf node 2-2 (i.e., the average of the sum of the actual shooting results corresponding to each training sample in the training sample subset 808-2) has a residual and needs to be fitted to meet the generation conditions of the next classification tree model. In the training process of the next classification tree model, leaf node 2-2 is used as the parent node of the next leaf node to complete the residual fitting. Alternatively, if both error 810-1 and error 810-2 do not exceed the threshold, it is considered that a set of classification tree models has been trained and there is no need to generate the next classification tree model.

[0142] The embodiments provided in this application set the node corresponding to the first shooting attribute feature as the root node in the first classification tree model to be generated; the training samples in the training sample set are classified according to the classification conditions corresponding to the first shooting attribute feature to obtain multiple training sample subsets; the node corresponding to each training sample subset in the multiple training sample subsets is set as a leaf node in the first classification tree model, wherein the leaf node in the first classification tree model is a child node of the root node in the first classification tree model, and the target value set on the leaf node in the first classification tree model is the average value of the actual shooting results in the corresponding training sample subset; based on the error between the target value set on the leaf node in the first classification tree model and the actual shooting results in the corresponding training sample subset, it is determined whether to generate a second classification tree model, and the error is used to determine whether to generate a next classification tree model, thereby achieving the purpose of fitting the residual of the previous classification tree model with the next classification tree model and improving the accuracy of model prediction.

[0143] As an optional approach, a second classification tree model group is generated using the node corresponding to the second shooting attribute feature as the root node, including:

[0144] S1, set the node corresponding to the second shooting attribute feature as the root node in the third classification tree model to be generated;

[0145] S2, classify the training samples in the training sample set according to the classification conditions corresponding to the second shooting attribute features to obtain multiple training sample subsets;

[0146] S3, set the node corresponding to each training sample subset in the multiple training sample subsets as a leaf node in the third classification tree model, where the leaf node in the third classification tree model is a child node of the root node in the third classification tree model, and the target value set on the leaf node in the third classification tree model is the average value of the residual between the actual shooting result in the corresponding training sample subset and the average value of the actual shooting result in the training sample subset.

[0147] S4. Based on the error between the target value and the residual value set on the leaf node of the third classification tree model, determine whether to generate a fourth classification tree model.

[0148] It should be noted that, unlike the above embodiments which determine whether to generate the next classification tree model, in this embodiment, the determination of whether to generate the next classification tree model is based on the error between the target value set on the leaf node of the third classification tree model and the residual value.

[0149] As an optional approach, based on the error between the target values ​​set on the leaf nodes of the first classification tree model and the actual shooting results in the corresponding training sample subset, it is determined whether to generate a second classification tree model, including:

[0150] When the error between the average value set at the leaf node of the first classification tree model and the actual shooting result in the corresponding training sample subset is greater than the third preset threshold, the second classification tree model is determined to be generated, and the second classification tree model is generated based on the first classification tree model.

[0151] As an optional approach, a second classification tree model is generated based on the first classification tree model, including:

[0152] S1, determine the first leaf node in the first classification tree model, wherein the error between the average value set on the first leaf node and the actual shooting result in the training sample subset corresponding to the first leaf node is greater than the third preset threshold.

[0153] S2, based on the feature value distribution of the shooting attribute features other than the first shooting attribute feature in the shooting attribute feature set in the training sample set, select the second shooting attribute feature in the shooting attribute feature set. Here, multiple training sample subsets are training sample subsets obtained by training samples in the training sample set according to the classification conditions corresponding to the second shooting attribute feature. The feature value of the second shooting attribute feature in each of the multiple training sample subsets corresponds to a feature value cluster.

[0154] S3, replace the first leaf node in the first classification tree model with the child node corresponding to the second shooting attribute feature, and set the node corresponding to each training sample subset in multiple training sample subsets as the leaf node of the child node corresponding to the second shooting attribute feature to obtain the second classification tree model.

[0155] It should be noted that, in the first classification tree model, the first leaf node is determined; based on the feature value distribution of the shooting attribute features other than the first shooting attribute feature in the shooting attribute feature set in the training sample set, the second shooting attribute feature is selected in the shooting attribute feature set; the first leaf node in the first classification tree model is replaced with the child node corresponding to the second shooting attribute feature, and the node corresponding to each training sample subset in multiple training sample subsets is set as the leaf node of the child node corresponding to the second shooting attribute feature, thus obtaining the second classification tree model.

[0156] To further illustrate, the optional based on Figure 8 The scenario shown continues, for example... Figure 9As shown, assuming the error 810-1 is less than the third threshold, while the error 810-2 is greater than the third threshold, it can be understood, but is not limited to, that leaf node 2-1 has residuals to be fitted, while leaf node 2-2 has no residuals and therefore does not need to be fitted. Figure 9 As shown in (a) in the middle;

[0157] Furthermore, such as Figure 9 As shown in (b), to fit the residual existing in leaf node 2-1, the child node 2-1 corresponding to leaf node 2-1 is used as the parent node of leaf node 3-1 and leaf node 3-2 in the newly generated classification tree model 902 (second classification tree model). Based on this, the first layer of the classification tree model 902 has a root node 1-1, and the root node 1-1 is set with classification condition 806; the second layer of the classification tree model 902 has a child node 2-2, and the child node 2-1 is set with classification condition 904. The second layer of the tree structure also has a leaf node 2-2, and the weight value of leaf node 2-2 is the target value corresponding to each training sample in the training sample subset 808-2; the third layer of the classification tree model 902 has leaf nodes 3-1 and 3-2, and the weight values ​​of leaf nodes 3-1 and 3-2 will be passed through the classification tree. Figure 8 The logic shown is used for calculation;

[0158] Furthermore, classification condition 904 may, but is not limited to, be the classification condition corresponding to the second shooting attribute feature selected in the shooting attribute feature set 802 based on the feature value distribution of shooting attribute features other than the first shooting attribute feature (the shooting attribute feature used by the classification tree model 804) in the shooting attribute feature set in the training sample set 802.

[0159] Through the embodiments provided in this application, a first leaf node is determined in the first classification tree model; based on the feature value distribution of the shooting attribute features other than the first shooting attribute feature in the shooting attribute feature set in the training sample set, a second shooting attribute feature is selected in the shooting attribute feature set; the first leaf node in the first classification tree model is replaced with the child node corresponding to the second shooting attribute feature, and the node corresponding to each training sample subset in multiple training sample subsets is set as the leaf node of the child node corresponding to the second shooting attribute feature, thereby obtaining the second classification tree model. This achieves the purpose of fitting the residuals existing in the previous classification tree model by generating the next classification tree model, and realizes the effect of improving the accuracy of model prediction.

[0160] As an optional approach, a set of classification tree models is generated using the node corresponding to the first shooting attribute feature as the root node, including:

[0161] S1, if the node depth in the i-th classification tree model is less than the fourth preset threshold, perform the following steps, where i is a natural number:

[0162] S2, when the error between the average value set on the leaf node of the i-th classification tree model and the actual shooting result in the corresponding training sample subset is greater than the third preset threshold, determine to generate the (i+1)-th classification tree model;

[0163] S3, determine the first leaf node in the i-th classification tree model, wherein the error between the average value set on the first leaf node and the actual shooting result in the training sample subset corresponding to the first leaf node is greater than the third preset threshold.

[0164] S4. Based on the feature value distribution of the shooting attribute features in the shooting attribute feature set in the training sample set, excluding the first shooting attribute feature to the i-th shooting ray feature, select the (i+1)-th shooting attribute feature in the shooting attribute feature set. Here, multiple training sample subsets are training sample subsets obtained by training samples in the training sample set according to the classification conditions corresponding to the (i+1)-th shooting attribute feature. The feature value of the (i+1)-th shooting attribute feature in each of the multiple training sample subsets corresponds to a feature value cluster.

[0165] S5, replace the first leaf node in the i-th classification tree model with the i+1-th child node corresponding to the i+1-th shooting attribute feature, and set the node corresponding to each training sample subset in the multiple training sample subsets as the leaf node of the i+1-th child node, thus obtaining the i+1-th classification tree model.

[0166] It should be noted that the training iterations of the classification tree model in the lateral conceptual dimension can be limited to a pre-set node depth, i.e., the number of tree levels, such as... Figure 9 As shown, if the node depth is preset to 3, it means that the maximum number of tree layers in the classification tree model is 3. Figure 9 In the classification tree model 902 shown in (b), the number of tree layers is 3, which means that the maximum number of tree layers corresponding to the maximum node depth has been reached. Therefore, it is not allowed to generate the next classification tree model, but instead, a set of classification tree models is directly output; otherwise, it continues to generate.

[0167] Furthermore, in this embodiment, the training iterations of the classification tree model in the lateral concept can also be, but are not limited to, a pre-set number of iterations, i.e., the total number of generated classification tree models, such as... Figure 9 As shown, if the number of iterations is set to 2, it means that the maximum number of generated classification tree models is 2. Figure 9The classification tree models 804 and 902 shown in the diagram have reached a total of 2, which means that the maximum number of classification tree models corresponding to the maximum number of iterations has been reached. Therefore, it is not allowed to generate the next classification tree model, but instead, a set of classification tree models is directly output; otherwise, it continues to generate.

[0168] Furthermore, in this embodiment, the training iteration of the classification tree model in the lateral concept can also be, but is not limited to, a pre-set first convergence condition. For example, the predicted shooting result output by each classification tree model is calculated and compared with the actual shooting prediction result. If the comparison result satisfies the first convergence condition, the next classification tree model will not be generated; otherwise, it will continue to be generated.

[0169] In the embodiments provided in this application, when the node depth in the i-th classification tree model is less than a fourth preset threshold, the following steps are performed, where i is a natural number: When the error between the average value set on the leaf node of the i-th classification tree model and the actual shooting result in the corresponding training sample subset is greater than a third preset threshold, the (i+1)-th classification tree model is determined; a first leaf node is determined in the i-th classification tree model, wherein the error between the average value set on the first leaf node and the actual shooting result in the training sample subset corresponding to the first leaf node is greater than a third preset threshold; based on the feature value distribution of the shooting attribute features in the shooting attribute feature set in the training sample set, excluding the first shooting attribute feature to the i-th shooting ray feature, in the shooting... The (i+1)th shooting attribute feature is selected from the attribute feature set. Multiple training sample subsets are training sample subsets obtained by classifying the training samples in the training sample set according to the classification conditions corresponding to the (i+1)th shooting attribute feature. The feature value of the (i+1)th shooting attribute feature in each of the multiple training sample subsets corresponds to a feature value cluster. The first leaf node in the i-th classification tree model is replaced with the (i+1)th child node corresponding to the (i+1)th shooting attribute feature, and the node corresponding to each of the multiple training sample subsets is set as the leaf node of the (i+1)th child node, resulting in the (i+1)th classification tree model. This achieves the goal of iterative training of the model in the horizontal dimension, ensuring the training quality of the model.

[0170] As an optional approach, after obtaining the (i+1)th classification tree model, the following steps are taken:

[0171] S1, in the (i+1)th classification tree model, obtain the error between the average value set on the leaf node of the (i+1)th child node and the actual shooting result in the training sample subset corresponding to the leaf node of the (i+1)th child node;

[0172] S2, when there is a leaf node in the (i+1)th child node whose error is greater than the third preset threshold, adjust the classification conditions corresponding to the (i+1)th shooting attribute feature set on the (i+1)th child node.

[0173] Optionally, in this embodiment, the training iteration of the classification tree model in the vertical concept may be, but is not limited to, the second convergence condition. For example, the predicted shooting result output by each classification tree model is calculated and compared with the actual shooting prediction result. If the comparison result satisfies the second convergence condition, the generation judgment of the next classification tree model is initiated; otherwise, the classification conditions or other values ​​of the current classification tree model are adjusted.

[0174] Through the embodiments provided in this application, in the (i+1)th classification tree model, the error between the average value set on the leaf node of the (i+1)th child node and the actual shooting result in the training sample subset corresponding to the leaf node of the (i+1)th child node is obtained; when there is a leaf node in the (i+1)th child node with an error greater than the third preset threshold, the classification conditions corresponding to the (i+1)th shooting attribute feature set on the (i+1)th child node are adjusted, thereby achieving the purpose of completing the iterative training of the model in the vertical dimension and ensuring the training quality of the model.

[0175] As an optional approach, the first shooting attribute feature is selected from the set of shooting attribute features, which includes: selecting N shooting attribute features from the set of shooting attribute features, where N is a natural number greater than 1;

[0176] As an optional approach, a set of classification tree models is generated using the node corresponding to the first shooting attribute feature as the root node. This includes: using the node corresponding to one of the N shooting attribute features as the root node, generating a set of classification tree models corresponding to each shooting attribute feature in parallel, resulting in a total of M sets of classification tree models. Here, a set of classification tree models includes M sets of classification tree models, where M is a natural number greater than or equal to N and greater than 1.

[0177] Optionally, in this embodiment, to improve the running efficiency of the classification tree model, multi-line classification tree models can be trained / run in parallel to obtain multiple classification tree model sets.

[0178] To further illustrate, the optional based on Figure 4 The scenario shown continues, for example... Figure 10 As shown, during the training of classification tree model 11, classification tree model 12... classification tree model 1n, classification tree model 31, classification tree model 32... classification tree model 3k will also be trained synchronously to achieve the effect of generating a set of classification tree models corresponding to each shooting attribute feature in parallel, so as to efficiently obtain M sets of classification tree models.

[0179] Optionally, the difference between training classification tree models 31, 32, ..., 3k and training classification tree models 11, 12, ..., 1n may be, but is not limited to, different classification conditions set at the root nodes. For example, during the training of classification tree models 31, 32, ..., 3k, the classification condition set at the root node is classification condition 1002, while during the training of classification tree models 11, 12, ..., 1n, the classification condition set at the root node is classification condition 402.

[0180] Through the embodiments provided in this application, each of the N shooting attribute features has its corresponding node as the root node, and a set of classification tree models corresponding to each shooting attribute feature is generated in parallel, resulting in a total of M sets of classification tree models. Each set of classification tree models includes M sets of classification tree models, where M is a natural number greater than or equal to N and greater than 1. This achieves the goal of generating a set of classification tree models corresponding to each shooting attribute feature in parallel, thereby improving the training efficiency of the classification tree models.

[0181] As an optional approach, each of the N shooting attribute features is used as the root node to generate a set of classification tree models corresponding to each shooting attribute feature in parallel, resulting in a total of M sets of classification tree models, including:

[0182] S1, taking the node corresponding to one of the N shooting attribute features as the root node, generates a set of classification tree models corresponding to each shooting attribute feature in parallel, resulting in a total of N sets of classification tree models; or,

[0183] S2, taking the node corresponding to one of the N shooting attribute features as the root node and the node corresponding to another or more of the N shooting attribute features as the child nodes of the root node, generates a set of classification tree models corresponding to each shooting attribute feature in parallel, resulting in a total of M sets of classification tree models.

[0184] As an optional approach, a set of classification tree models is generated based on the training sample set, including:

[0185] S1, Select the current shooting attribute feature from the set of shooting attribute features;

[0186] S2, using the node corresponding to the current shooting attribute feature as the root node and the current classification condition corresponding to the current shooting attribute feature as the classification condition corresponding to the root node, generate the current classification tree model. The leaf nodes in the current classification tree model group are set with the current value, which includes the result average or the residual average.

[0187] S3, if the error corresponding to the current value has not reached the convergence condition, select the next shooting attribute feature from the shooting attribute feature set as the current shooting attribute feature, or, use the next classification condition corresponding to the current shooting attribute feature as the current classification condition, or, select the next shooting attribute feature from the shooting attribute feature set as the current shooting attribute feature and use the classification condition corresponding to the next shooting attribute feature as the current classification condition.

[0188] S4. If the error corresponding to the current value reaches the convergence condition, determine a set of classification tree models including the current classification tree model.

[0189] Optionally, in this embodiment, since the shooting attribute feature set includes multiple different shooting attribute features, and each shooting attribute feature can constitute multiple different classification conditions in different value ranges, it is assumed that the shooting attribute feature set includes N different shooting attribute features, and each shooting attribute feature can constitute M different classification conditions in different value ranges. Then, there are a total of N×M cases that can be used as the classification conditions corresponding to the root node.

[0190] Optionally, the optimal combination of shooting attribute features and value ranges from the aforementioned N×M cases can be selected as the configuration of the root node for the set of classification tree models to be generated. The specific selection method may include, but is not limited to, setting the shooting attribute features and value ranges corresponding to each of the aforementioned N×M cases as the root node in sequence, and inputting them into the training samples in the training sample set to obtain the corresponding output value for each case. Then, the output value is compared with the actual shooting result of each training sample to obtain the corresponding error for each case. Based on the error, the combination of shooting attribute features and value ranges with the smallest error from the aforementioned N×M cases can be selected as the configuration of the root node for the set of classification tree models to be generated.

[0191] Furthermore, in this embodiment, to improve model training efficiency, the shooting attribute features and their value ranges corresponding to each of the above N×M cases can be set as the root node in parallel, but are not limited to.

[0192] As an optional approach, obtain the training sample set corresponding to the set of virtual shooting props, including:

[0193] S1, obtain the configuration file and corresponding actual shooting results for each virtual shooting item in the virtual shooting item set. The configuration file includes a set of shooting attribute features, and the actual shooting results include battle data using a virtual shooting item.

[0194] S2 generates a training sample set based on the configuration file of each virtual shooting prop and the corresponding actual shooting results.

[0195] Optionally, in this embodiment, the combat data of a virtual shooting prop can be, but is not limited to, the KD result, i.e., the kill-to-death ratio.

[0196] It's important to note that traditional testing methods for weapon balance require running numerous test matches against similar weapons, then analyzing K / D ratios to determine weapon strength. Another approach involves pulling data from external websites on already released games, such as usage frequency and K / D ratios, and analyzing win rate data to further determine weapon strength. However, existing methods require significant manpower for testing weapon balance, conducting match tests, and observing and comparing weapon K / D results, which is extremely labor-intensive. Furthermore, match results are easily influenced by tester performance, leading to low accuracy. The other approach requires analysis of external data after the project is launched, resulting in significant delays. Additionally, it cannot provide timely assessments when new weapons need to be evaluated before release.

[0197] In this embodiment, the attributes of firearms can be parsed from the latest planned firearm configuration list, and predicted using machine learning. The KD value of the firearm can then be directly output on the corresponding page. After launch, the KD of the firearm in the game can be retrieved periodically and displayed on the corresponding platform. The error between the predicted KD and the external network KD can be compared, and this error can be used to build a prediction model. Furthermore, if the predicted KD exceeds a normal threshold, the predicted KD can be lowered by adjusting (weakening) the firearm's attributes to maintain game balance. Conversely, if the predicted KD is below a normal threshold, the predicted KD can be increased by adjusting (strengthening) the firearm's attributes, also contributing to game balance.

[0198] To further illustrate, the model is first trained on historical weapon combat data using AI algorithms to build a prediction model. Then, target weapon feature data is extracted as the sample to be predicted. The model is then used to predict the expected KD value of the sample. Optionally, the weapon feature data can be extracted by parsing the weapon configuration file on the client side. The feature data quantification indicators are filtered, and there are a total of 296 items: headshot damage multiplier, torso damage multiplier, firing interval (the time interval between two shots fired in burst mode), vertical recoil (base value, modification factor, maximum value), horizontal recoil (base value, modification factor, maximum value), recoil time (time required to raise and lower the weapon), and weapon raising time, etc. Therefore, a training sample consists of a 297-dimensional vector (296 features + 1 label). The model is then built upon this vector. Historical weapon combat KD data can be stored in the TDW data warehouse and stored in a local MySQL database, with the data format being KD information for different ranks in the game.

[0199] Furthermore, adding optional sample features can negatively impact model efficiency, and maintaining a large number of features also presents efficiency challenges. Therefore, it's necessary to filter sample features, removing those with lower importance. Reducing the number of features leads to a decrease in model accuracy; thus, the degree of filtering for effective features needs to be balanced between accuracy and efficiency.

[0200] To further illustrate, the training process of the KD model can be optionally predicted as follows: Figure 11 As shown, the KD value data of firearms (KD result data 1102) is obtained from the external network and stored in the database. Based on the weapon attribute configuration table 1102 and the firearm configuration file in the configuration attribute table 1104, the firearm attribute configuration table 1106 is constructed. Then, based on the firearm attribute configuration table 1106 and the KD result data 1102, a vector composed of firearm attribute features and KD values ​​is generated and used as training sample 1110.

[0201] Specifically, the features are preprocessed, mainly by filling in missing values, filling in the mean, and processing discrete features. After further preprocessing, the firearm feature samples are formed as I = {(x1, y1), (x2, y2), ... (xn, yn)}, where x and y represent the feature vector and the label kd value, respectively, and n is the number of samples. The samples can be randomly split into training and prediction sets in a 7:3 ratio to improve the utilization efficiency of the training samples.

[0202] In addition, the model parameters of the prediction model 1112 can be set, such as the maximum number of iterations Boost_Num, the number of base learners Eastimators, the learning rate ETA, the evaluation function MAE, the splitting index Gamma, the maximum depth Max_Depth, etc. Finally, the training is repeated for multiple rounds, and the model with the smallest error is selected as the final prediction model 1112 and saved as output.

[0203] To further illustrate, the optional prediction of the KD model's operation process can be based on... Figure 11 The scene shown Figure 12 As shown, the staff updates the firearm configuration table and generates feature samples of various attachments for the new firearms based on the new configuration table (without KD value labels, including new firearm data 1212, and may include but is not limited to accessory data 1214). The feature samples are then input into the trained prediction model 1212 to predict the predicted KD results 1216 for various attachments of the new firearms.

[0204] Alternatively, as an alternative implementation method, such as Figure 13 As shown, the methods for predicting shooting results include:

[0205] S1302, Obtain the target shooting attribute feature set of the virtual shooting prop to be predicted;

[0206] S1304. Based on the shooting attribute features and corresponding classification conditions set on the non-leaf nodes in the target classification tree model, and the target shooting attribute feature set, determine the target leaf node corresponding to the target virtual shooting prop in the leaf nodes of the target classification tree model. The target shooting attribute feature set satisfies the classification conditions set on the parent node of the target leaf node.

[0207] S1306, Obtain the target value set on the target leaf node, wherein the target value includes the average value of the actual shooting results and the average residual of the difference between the actual shooting results and the average value of the results;

[0208] S1308, Based on the target value set on the target leaf node, determine the predicted shooting result of the target virtual shooting prop;

[0209] The target classification tree model is selected from a set of classification tree models based on the target values ​​set on the leaf nodes and the actual shooting results in the corresponding training sample subset. The set of classification tree models is generated based on the training sample set. Each training sample in the training sample set includes a set of shooting attribute features of a virtual shooting prop in the set of virtual shooting props and the actual shooting result obtained by using a virtual shooting prop. The set of shooting attribute features includes shooting attribute features in multiple dimensions. The training sample set includes a subset of training samples.

[0210] Optionally, in this embodiment, the above-mentioned method for predicting shooting results can be applied, but is not limited to, in scenarios where the battle balance information of virtual shooting props is tested (used to measure whether the virtual shooting props disrupt the game balance of the virtual game in which the virtual shooting props are located due to excessively high / low prop strength). For example, a training sample set composed of multiple virtual shooting props as samples is first obtained, and each of the multiple virtual shooting props as samples carries historical usage data or is directly marked with one or more actual shooting results; then, the training sample set is used to train a classification tree model, and a set of classification tree models is obtained through training; furthermore, to ensure the accuracy of the model output, a target classification tree model is selected from the set of classification tree models as the actual prediction model;

[0211] Based on this, the item data (such as item type, item model, item accessories, etc.) of the virtual shooting item to be tested is input into the target classification tree model, and the output of the target classification tree model is used as the predicted shooting result of the virtual shooting item to be tested. Then, based on the obtained predicted shooting result, the balance of the virtual shooting item to be predicted is judged (the judgment result can be understood as balance information, but is not limited to). Based on the judgment result, the item data of the virtual shooting item to be predicted is adjusted (or not adjusted) to ensure that the virtual shooting item to be predicted will not disrupt the game balance of the virtual game it belongs to.

[0212] To further illustrate, alternatively, for example... Figure 3As shown, assuming the virtual shooting prop to be predicted includes the following prop data: weapon type is rifle, model 25, accessories include a rear grip of model 1, a skill (GunPerk) of model 3, a scope of model 3, a grip of model 2, a muzzle of model 2, and a stock of model 1, then the above prop data is entered into the shooting result prediction interface 302. The shooting result prediction interface 302 also has virtual buttons "Get" and "Predict". The virtual button "Get" is used to trigger the acquisition of existing predicted shooting results (if the current virtual shooting prop to be predicted is being predicted for the first time, or there are no existing predicted shooting results, the virtual button "Get" cannot be triggered, or after triggering, a prompt message indicating that there are no existing predicted shooting results will be displayed). The virtual button "Predict" is used to trigger the prediction command to perform a prediction operation on the current virtual shooting prop to be predicted, such as... Figure 3 As shown in (a), selecting the virtual button "Predict" has triggered the prediction command;

[0213] Furthermore, such as Figure 3 As shown in (b), in response to the prediction command, the predicted shooting result of the target virtual shooting prop is displayed in the prediction shooting result interface 304 (e.g., "0.9219936"). In addition, the prediction shooting result interface 304 is also equipped with a virtual button "Approve" and a virtual button "Cancel". The virtual button "Approve" is used to trigger the storage of the currently predicted shooting result in the existing prediction shooting results, and the virtual button "Cancel" is used to trigger the deletion of the currently predicted shooting result.

[0214] Optionally, in virtual shooting games where virtual shooting props are located, there are often many different types of virtual shooting prop main bodies, and each virtual shooting prop main body can be used as a virtual shooting prop to test or predict shooting results. In addition, virtual shooting games often have many different types of virtual shooting prop accessories, and virtual shooting prop main bodies and virtual shooting prop accessories can be flexibly combined to form new virtual shooting props. This means that testing or predicting shooting results requires not only testing all types of virtual shooting prop main bodies, but also testing new virtual shooting props obtained through any combination method, which dramatically increases the number of objects to be predicted or tested, and the pressure of prediction or testing also increases.

[0215] Optionally, in this embodiment, whether it is a virtual shooting prop body that is used alone as a virtual shooting prop, or a virtual shooting prop body and virtual shooting prop accessories that are used in combination as virtual shooting props, they can be regarded as prop data with the same or different shooting attribute characteristics, without distinguishing the type or combination method. Therefore, even if there are a large number of objects to be predicted or tested, there is no prediction or testing pressure in this embodiment.

[0216] Optionally, in this embodiment, the shooting attribute features may be, but are not limited to, attribute features in the shooting game scene that are highly correlated with the actual shooting results, such as the damage multiplier of a virtual bullet hitting the head of an enemy virtual object, the damage multiplier of a virtual bullet hitting the torso of an enemy virtual object, the firing interval (such as the time interval between two bullets firing in burst mode), the vertical recoil of the shot (base value, modification factor, maximum value), the horizontal recoil of the shot (base value, modification factor, maximum value), the time to put away the virtual shooting prop (the time required to raise the gun and put it down), and the time to raise the virtual shooting prop, etc.

[0217] Optionally, in this embodiment, the actual shooting result can be used, but is not limited to, to represent the combat result fed back during one or more rounds of shooting games using virtual shooting props, such as kill result, kill result, assist result, etc. Furthermore, predicting the shooting result of a virtual shooting prop can be understood, but is not limited to, predicting the combat result that the virtual shooting prop should feed back in each future round of shooting games.

[0218] Optionally, in this embodiment, the classification tree model may include, but is not limited to, a root node (non-leaf node), one or more leaf nodes, and one or more child nodes (non-leaf nodes). The root node can be understood as an input node, that is, classifying the shooting attribute features input to the classification tree model and outputting the classification result to the next node (which can be a child node or a leaf node). If the next node is a corresponding leaf node, a preset value of the corresponding leaf node is output, and the output result of the current classification tree model is determined based on this preset data. Optionally, in the classification tree model, each non-leaf node may have, but is not limited to, a preset classification condition, and this corresponding classification condition is used to classify the input features for output to the next node.

[0219] Optionally, in this embodiment, a set of classification tree models includes multiple classification tree models, and the target classification tree model selected from the set of classification tree models may include, but is not limited to, multiple classification tree models. In other words, the selection criteria in a set of classification tree models may include, but is not limited to, selecting all classification tree models that meet the preset conditions as the target classification tree model; furthermore, the target classification tree model selected from a set of classification tree models may include, but is not limited to, a single classification tree model. In other words, the selection criteria in a set of classification tree models may include, but is not limited to, selecting the classification tree model that best meets the preset conditions as the target classification tree model.

[0220] It should be noted that for ordinary prediction scenarios, model technology is undoubtedly the best way to ensure the accuracy and efficiency of prediction. However, the prediction scenarios of actual shooting results are different from ordinary prediction scenarios. The prediction scenarios of actual shooting results have high scenario complexity (there are many and complex factors affecting the actual shooting results, such as the rate of fire will change at different ranges, and the damage caused by the shooting will also change depending on the hit site). Therefore, how to apply model technology to the prediction scenarios of actual shooting results is still a difficult problem that related technologies have not yet overcome.

[0221] In this embodiment, to overcome this challenge, a classification tree model in model technology is proposed to predict actual shooting results. Furthermore, by setting shooting attribute features and corresponding classification conditions from the shooting attribute feature set on the non-leaf nodes of the classification tree model, shooting attribute features and corresponding classification conditions that closely match the shooting result prediction scenario are combined into the classification tree model, resulting in a set of classification tree models with a high degree of fit with the shooting result prediction scenario. The training process of the classification tree model is further improved by setting the average value of actual shooting results from the training sample subset on the leaf nodes. Finally, a target classification tree model is selected from the set of classification tree models for actual shooting result prediction, ensuring the accuracy of the shooting result prediction output by the model.

[0222] Further examples, such as Figure 4 As shown, firstly, the root node 1-1 in the classification tree model 11 is constructed using all or part of the shooting attribute features in the shooting attribute feature set of the training sample set 404, and the classification condition 402 set on the root node 1-1 is constructed. From this, two leaf nodes with the root node 1-1 as the parent node can be obtained. Then, each training sample in the training sample set is input into the root node 1-1 in sequence, and the classification condition 402 is used to classify each input training sample. The classification result is to be classified into leaf node 1-2 or leaf node 1-3.

[0223] After classifying each training sample using classification condition 402, the training sample subsets containing leaf nodes 1-2 and 1-3 are statistically analyzed, and the average value (average result) of the actual shooting results recorded in the training sample set 404 for each training sample subset is calculated. This average value is then used as the weight value of the corresponding leaf node.

[0224] Furthermore, optionally, after obtaining the average value of the actual shooting results (sum) corresponding to each training sample subset, the average value can be compared with each actual shooting result recorded in the training sample set 404 for each training sample subset to obtain the corresponding error. If the error does not meet the preset conditions, the weight values ​​of classification condition 402 and / or leaf nodes can be adjusted. That is, the determined weight values ​​of classification condition 402 and / or leaf nodes can be, but are not limited to, adjustable; and the weight values ​​of classification condition 402 and / or leaf nodes can be set to non-adjustable after the classification tree model 11 has been trained.

[0225] Furthermore, after the classification tree model 11 has been trained, it is possible, but not limited to, to continue generating the next classification tree model, namely classification tree model 12. The generation of classification tree model 12 is based on the root node 1-1 and leaf node 1-3 of classification tree model 11, and then uses all or part of the shooting attribute features in the shooting attribute feature set to construct child node 1-2 and classification condition 404, thereby obtaining two leaf nodes with child node 1-2 as the parent node, namely leaf node 1-4 and leaf node 1-5;

[0226] Optionally, in the process of generating classification tree model 12 based on classification tree model 11, the child nodes corresponding to the node positions of leaf nodes 1-2 and / or leaf nodes 1-3 can be selected as the parent nodes of the leaf nodes of classification tree model 12, and the selection criteria can be, but are not limited to, the errors corresponding to leaf nodes 1-2 and leaf nodes 1-3. For example, the child nodes corresponding to the node positions of leaf nodes with larger errors can be selected as the parent nodes of the leaf nodes of classification tree model 12; or, the child nodes corresponding to the node positions of leaf nodes with errors exceeding the error threshold can be selected as the parent nodes of the leaf nodes of classification tree model 12.

[0227] Furthermore, based on the same logic, the nth classification tree model, i.e., classification tree model 1n, is generated. Among the n classification tree models, the one with the smallest error is selected as the first target tree model, which is used to generate the initial predicted shooting result of the target virtual shooting prop. Although the error of the first target tree model is relatively the smallest, residuals inevitably still exist. Therefore, it is necessary to use the average result corresponding to each leaf node in the first target tree model to calculate the difference with the actual shooting result corresponding to each leaf node to obtain multiple differences. Then, these multiple differences are summed and averaged to obtain the average residual.

[0228] Optionally, in this embodiment, based on Figure 4 The scenario shown continues, for example... Figure 5As shown, assuming that the classification tree model 2 has been determined as the first target tree model 502, the training sample subset corresponding to each leaf node is obtained, and the actual shooting result of the training sample subset corresponding to each leaf node is modified to the difference between the average result and the actual shooting result. Alternatively, each training sample recorded in the original training sample set 404 records the actual shooting result as the sample label, but in this embodiment, after determining the first target classification tree model 502, the difference between the average result and the actual shooting result is used as the sample label for each training sample, thereby obtaining a new training sample set 506.

[0229] Furthermore, the root node 2-1 in the classification tree model 21 is constructed using all or part of the shooting attribute features in the shooting attribute feature set of the training sample set 506, along with the classification condition 502 set on the root node 2-1. Thus, two leaf nodes with the root node 2-1 as the parent node can be obtained. Then, each training sample in the training sample set 502 is sequentially input into the root node 2-1, and the classification condition 502 is used to classify each input training sample. The classification result is assigned to the leaf node 2-2 or the leaf node 2-3.

[0230] After classifying each training sample using classification condition 502, the training sample subsets containing leaf nodes 2-2 and 2-3 are statistically analyzed, and the average difference between the average result recorded in training sample set 502 and the actual shooting result (residual average) of each training sample subset is calculated. Then, the residual average is used as the weight value of the corresponding leaf node.

[0231] Furthermore, optionally, after obtaining the residual average value corresponding to each training sample subset, the residual average value can be compared with the difference between the average value of each result recorded in the training sample set 502 and the actual shooting result for each training sample subset to obtain the corresponding error. If the error does not meet the preset conditions, the weight values ​​of classification condition 502 and / or leaf nodes can be adjusted. That is, the determined weight values ​​of classification condition 502 and / or leaf nodes can be, but are not limited to, adjustable; and the weight values ​​of classification condition 502 and / or leaf nodes can be, but are not limited to, set to non-adjustable after the classification tree model 21 has been trained.

[0232] Furthermore, after the classification tree model 21 has been trained, it is possible, but not limited to, to continue generating the next classification tree model, namely classification tree model 22. The generation of classification tree model 22 is based on the root node 2-1 and leaf node 23 of classification tree model 21, and then uses all or part of the shooting attribute features in the shooting attribute feature set to construct child node 2-2 and classification condition 504, and thus obtains two leaf nodes with child node 2-2 as the parent node, namely leaf node 2-4 and leaf node 2-5;

[0233] Optionally, in the process of generating classification tree model 22 based on classification tree model 21, the child nodes corresponding to the node positions of leaf node 2-2 and / or leaf node 2-3 can be selected as the parent nodes of the leaf nodes of classification tree model 22, and the selection criteria can be, but are not limited to, the errors corresponding to leaf node 2-2 and leaf node 2-3. For example, the child nodes corresponding to the node positions of leaf nodes with larger errors can be selected as the parent nodes of the leaf nodes of classification tree model 22; or, the child nodes corresponding to the node positions of leaf nodes with errors exceeding the error threshold can be selected as the parent nodes of the leaf nodes of classification tree model 22.

[0234] Furthermore, based on Figure 4 The same logic in the illustrated embodiment generates the qth classification tree model, namely classification tree model 2q, and selects the classification tree model with the smallest error among the q classification tree models as the second target tree model, which is used to fit the residual existing in the first target tree model 502.

[0235] Furthermore, assuming that classification tree model 2q is determined as the second target classification tree model 510, then classification tree model 12 (first target classification tree model 502) and classification tree model 2q (second target classification tree model 510) can be used together as the target classification tree model 510, but not limited to.

[0236] Optionally, in this embodiment, a set of classification tree models may include, but is not limited to, including... Figure 4 The n classification tree models trained using a training sample set of 404 can also include, but are not limited to, the following: Figure 5The training sample set consists of 506 training trees, resulting in q classification tree models. Furthermore, the number of classification tree models in a set, n, q, or (n+q), can be determined by one or more constraints, including but not limited to. For example, the maximum number of iterations for each classification tree model can be pre-limited, for instance, set to n, meaning at most n trees will be generated during training, i.e., n iterations. Another example is pre-limiting the maximum tree depth, for instance, setting it to 10 layers, meaning the depth of each tree cannot exceed 10 layers during training. Yet another example is pre-setting convergence conditions, such that if any tree's output satisfies these convergence conditions during training, training ends, or in other words, the currently generated classification tree model is determined to be a set of classification tree models.

[0237] To further illustrate, the optional assumptions have already been... Figure 5 If a target classification tree model is selected from the set of classification tree models shown, then the process of generating the predicted shooting result of the target virtual shooting prop can be, but is not limited to, as follows: Figure 6 As shown, firstly, the shooting item data 602 of the target virtual shooting item 604 is obtained, and then the shooting item data 602 of the target virtual shooting item 604 is input into the target classification tree model 606; the target classification tree model 606 is as follows. Figure 6 As shown, the tree structure includes three layers. The first layer includes a root node 1-1, which has a classification condition 606-1. The second layer includes a child node 2-1 and leaf nodes 1-6, where the child node 2-1 has a classification condition 606-2, and the weight of leaf node 1-6 is assumed to be 0.6. The third layer includes a leaf node 3-1 and a leaf node 3-2, with the weight of leaf node 3-1 assumed to be 0.3 and the weight of leaf node 3-2 assumed to be 0.2.

[0238] Furthermore, based on the matching degree between shooting prop data 602 and classification condition 606-1, the target virtual shooting prop 604 is classified at the root node 1-1. Assuming the target virtual shooting prop 604 is classified to leaf node 1-6, the output predicted shooting result 608 is the weight value of leaf node 1-6, which is 0.6. Conversely, if the target virtual shooting prop 604 is classified to child node 2-1, then based on the matching degree between shooting prop data 602 and classification condition 606-2, the target virtual shooting prop 604 is classified at child node 2-1. If the target virtual shooting prop 604 is classified to leaf node 3-1, the output predicted shooting result 608 is the weight value of leaf node 3-1, which is 0.3 (target value); or if the target virtual shooting prop 604 is classified to leaf node 3-2, the output predicted shooting result 608 is the weight value of leaf node 3-2, which is 0.2.

[0239] Furthermore, in this embodiment, it is also possible, but not limited to, the following: Figure 7 As shown, firstly, the shooting item data 702 of the target virtual shooting item 704 is obtained, and then the shooting item data 702 of the target virtual shooting item 704 is input into the target classification tree model (including the first target classification model 706 and the second target classification model 708); wherein, the first target classification model 706 is as follows Figure 7 As shown, the tree structure includes three layers. The first layer includes a root node 1-1, which has a classification condition 706-1. The second layer includes a child node 1-1 and a leaf node 1-2, where the child node 1-1 has a classification condition 706-2, and the weight of the leaf node 1-2 is assumed to be 0.7. The third layer includes a leaf node 1-3 and a leaf node 1-3, with the weights of the leaf nodes 1-3 and 1-3 respectively being 0.3 and 0.2.

[0240] Furthermore, the second target classification model 708, such as Figure 7 As shown, there are two tree structures. The first tree structure includes a root node 2-1, where the root node 2-1 is set with classification condition 708-1. The second tree structure includes leaf node 2-1 and leaf node 2-2, where it is assumed that the weight of leaf node 2-1 is -0.2 and the weight of leaf node 2-2 is -0.1.

[0241] Furthermore, based on the matching degree between the shooting prop data 702 and the classification condition 707-1, the target virtual shooting prop 704 is classified at root node 1-1 and root node 2-1. Assuming that the target virtual shooting prop 704 is classified to leaf node 1-2 and leaf node 2-1, the output predicted shooting result 708 is the sum of the weight value of leaf node 1-2 (0.7) and the weight value of leaf node 2-1 (-0.2), which is 0.5. Conversely, if the target virtual shooting prop 704 is classified to child node 1-1, then based on the matching degree between the shooting prop data 702 and the classification condition 707-2, the target virtual shooting prop 704 is classified at child node 2-1. If the target virtual shooting prop 704 is classified to leaf node 1-3 and leaf node 2-2, the output predicted shooting result 708 is the sum of the weight value of leaf node 1-3 (0.3) and the weight value of leaf node 2-2 (-0.1), which is 0.2.

[0242] The embodiments provided in this application obtain a set of target shooting attribute features for the virtual shooting prop to be predicted; according to the shooting attribute features set on the non-leaf nodes of the target classification tree model and the corresponding classification conditions, and the set of target shooting attribute features, the target leaf node corresponding to the virtual shooting prop is determined in the leaf nodes of the target classification tree model, wherein the set of target shooting attribute features satisfies the classification conditions set on the parent node of the target leaf node; the target value set on the target leaf node is obtained, wherein the target value includes the average value of the actual shooting result and the average residual of the difference between the actual shooting result and the average value of the result; based on the target value set on the target leaf node, the predicted shooting result of the virtual shooting prop is determined. This provides an efficient means of predicting actual shooting results through model technology, and also ensures the accuracy of the actual shooting result prediction by combining the classification tree model with the shooting result prediction scenario with high fit. Thus, it achieves the technical objective of improving the prediction efficiency of actual shooting results while ensuring the prediction accuracy of actual shooting results, thereby realizing the technical effect of providing an efficient and accurate means of predicting actual shooting results.

[0243] As an optional approach, obtaining the target value set on the target leaf node includes: when the target classification tree model includes P classification tree models, obtaining the average value of the results set on the target leaf node in the P classification tree models and the average value of the residuals, where P is a natural number greater than 1;

[0244] As an optional approach, the predicted firing result of the target virtual firing prop is determined based on the target value set on the target leaf node, including: determining the predicted firing result of the target virtual firing prop as equal to the sum of the result average and the residual average.

[0245] Optionally, in this embodiment, when the target classification tree model includes multiple classification tree models, since each classification tree model has at least one target leaf node, the number of actual shooting results obtained should be at least greater than or equal to the number of classification tree models. For example, if the target classification tree model includes 3 classification tree models, and each classification tree model has 2 target leaf nodes, the number of actual shooting results obtained can be, but is not limited to, 6 (2×3). In order to obtain a unique output value, it is also necessary to perform summation and averaging operations on the 6 actual shooting results to obtain the final predicted shooting result.

[0246] Through the embodiments provided in this application, when the target classification tree model includes P classification tree models, the average result value and the average residual value set on the target leaf nodes in the P classification tree models are obtained; the predicted shooting result of the target virtual shooting prop is determined to be equal to the sum of the average result value and the average residual value, thereby achieving the purpose of improving the calculation granularity in the prediction process and realizing the effect of improving the accuracy of the predicted shooting result.

[0247] As an optional approach, based on the shooting attribute features set on the non-leaf nodes of the target classification tree model, the corresponding classification conditions, and the target shooting attribute feature set, the target leaf node corresponding to the virtual shooting prop is determined in the leaf nodes of the target classification tree model, including:

[0248] Starting from the root node in the target classification tree model, search for nodes in the target classification tree model that satisfy the classification conditions of the target shooting attribute feature set, until the target leaf node is found.

[0249] To further illustrate, the optional assumptions have already been... Figure 5 If a target classification tree model is selected from the set of classification tree models shown, then the process of generating the predicted shooting result of the target virtual shooting prop can be, but is not limited to, as follows: Figure 6 As shown, firstly, the shooting item data 602 of the target virtual shooting item 604 is obtained, and then the shooting item data 602 of the target virtual shooting item 604 is input into the target classification tree model 606; the target classification tree model 606 is as follows. Figure 6 As shown, the tree structure includes three layers. The first layer includes a root node 1-1, which has a classification condition 606-1. The second layer includes a child node 2-1 and leaf nodes 1-6, where the child node 2-1 has a classification condition 606-2, and the weight of leaf node 1-6 is assumed to be 0.6. The third layer includes a leaf node 3-1 and a leaf node 3-2, with the weight of leaf node 3-1 assumed to be 0.3 and the weight of leaf node 3-2 assumed to be 0.2.

[0250] Furthermore, based on the matching degree between shooting prop data 602 and classification condition 606-1, the target virtual shooting prop 604 is classified at the root node 1-1. Assuming the target virtual shooting prop 604 is classified to leaf node 1-6, the output predicted shooting result 608 is the weight value of leaf node 1-6, which is 0.6. Conversely, if the target virtual shooting prop 604 is classified to child node 2-1, then based on the matching degree between shooting prop data 602 and classification condition 606-2, the target virtual shooting prop 604 is classified at child node 2-1. If the target virtual shooting prop 604 is classified to leaf node 3-1, the output predicted shooting result 608 is the weight value of leaf node 3-1, which is 0.3 (target value); or if the target virtual shooting prop 604 is classified to leaf node 3-2, the output predicted shooting result 608 is the weight value of leaf node 3-2, which is 0.2.

[0251] Furthermore, in this embodiment, it is also possible, but not limited to, the following: Figure 7 As shown, firstly, the shooting item data 702 of the target virtual shooting item 704 is obtained, and then the shooting item data 702 of the target virtual shooting item 704 is input into the target classification tree model (including the first target classification model 706 and the second target classification model 708); wherein, the first target classification model 706 is as follows Figure 7 As shown, the tree structure includes three layers. The first layer includes a root node 1-1, which has a classification condition 706-1. The second layer includes a child node 1-1 and a leaf node 1-2, where the child node 1-1 has a classification condition 706-2, and the weight of the leaf node 1-2 is assumed to be 0.7. The third layer includes a leaf node 1-3 and a leaf node 1-3, with the weights of the leaf nodes 1-3 and 1-3 respectively being 0.3 and 0.2.

[0252] Furthermore, the second target classification model 708, such as Figure 7 As shown, there are two tree structures. The first tree structure includes a root node 2-1, where the root node 2-1 is set with classification condition 708-1. The second tree structure includes leaf node 2-1 and leaf node 2-2, where it is assumed that the weight of leaf node 2-1 is -0.2 and the weight of leaf node 2-2 is -0.1.

[0253] Furthermore, based on the matching degree between the shooting prop data 702 and the classification condition 707-1, the target virtual shooting prop 704 is classified at root node 1-1 and root node 2-1. Assuming that the target virtual shooting prop 704 is classified to leaf node 1-2 and leaf node 2-1, the output predicted shooting result 708 is the sum of the weight value of leaf node 1-2 (0.7) and the weight value of leaf node 2-1 (-0.2), which is 0.5. Conversely, if the target virtual shooting prop 704 is classified to child node 1-1, then based on the matching degree between the shooting prop data 702 and the classification condition 707-2, the target virtual shooting prop 704 is classified at child node 2-1. If the target virtual shooting prop 704 is classified to leaf node 1-3 and leaf node 2-2, the output predicted shooting result 708 is the sum of the weight value of leaf node 1-3 (0.3) and the weight value of leaf node 2-2 (-0.1), which is 0.2.

[0254] The embodiments provided in this application start from the root node in the target classification tree model, search for nodes in the target classification tree model that satisfy the classification conditions of the target shooting attribute feature set, until the target leaf node is found. This provides an efficient means of predicting actual shooting results through model technology. Furthermore, by combining the classification tree model with the shooting result prediction scenario with a high degree of fit, the accuracy of the actual shooting result prediction is also guaranteed. Thus, the technical objective of improving the prediction efficiency of actual shooting results while ensuring the prediction accuracy of actual shooting results is achieved, thereby realizing the technical effect of providing an efficient and accurate means of predicting actual shooting results.

[0255] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0256] According to another aspect of the present invention, a training apparatus for a shooting result prediction model for implementing the training method of the above-described shooting result prediction model is also provided. For example... Figure 14 As shown, the device includes:

[0257] The first acquisition unit 1402 is used to acquire a training sample set corresponding to the set of virtual shooting props. Each training sample in the training sample set includes a shooting attribute feature set of a virtual shooting prop in the set of virtual shooting props and the actual shooting result obtained by using a virtual shooting prop. The shooting attribute feature set includes shooting attribute features of multiple dimensions.

[0258] The generation unit 1404 is used to generate a set of classification tree models based on the training sample set. The non-leaf nodes of the classification tree model are set with shooting attribute features from the shooting attribute feature set and corresponding classification conditions. The leaf nodes of the classification tree model are set with the target values ​​of the actual shooting results in the training sample subset. The target values ​​include the average value of the actual shooting results in the training sample subset and the average residual value of the difference between the actual shooting results and the average value of the results in the training sample subset. The shooting attribute feature set of each virtual shooting prop in the training sample subset satisfies the classification conditions corresponding to the parent node of the leaf node. The training sample set includes the training sample subset.

[0259] Selection unit 1406 is used to select a target classification tree model from a set of classification tree models based on the target values ​​set on the leaf nodes in the classification tree model and the actual shooting results in the training sample subset. The target classification tree model is used to generate the predicted shooting results of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes in the target classification tree model and the corresponding classification conditions, as well as the target values ​​set on the leaf nodes.

[0260] Optionally, in this embodiment, the training device for the shooting result prediction model can be applied, but is not limited to, in scenarios where the battle balance information of virtual shooting props is tested (used to measure whether the virtual shooting props disrupt the game balance of the virtual game in which the virtual shooting props are located due to excessively high / low prop strength). For example, a training sample set composed of multiple virtual shooting props as samples is first obtained, and each of the multiple virtual shooting props as samples carries historical usage data or is directly marked with one or more actual shooting results; then, the training sample set is used to train a classification tree model, and a set of classification tree models is obtained through training; furthermore, to ensure the accuracy of the model output, a target classification tree model is selected from the set of classification tree models as the actual prediction model;

[0261] Based on this, the item data (such as item type, item model, item accessories, etc.) of the virtual shooting item to be tested is input into the target classification tree model, and the output of the target classification tree model is used as the predicted shooting result of the virtual shooting item to be tested. Then, based on the obtained predicted shooting result, the balance of the virtual shooting item to be predicted is judged (the judgment result can be understood as balance information, but is not limited to). Based on the judgment result, the item data of the virtual shooting item to be predicted is adjusted (or not adjusted) to ensure that the virtual shooting item to be predicted will not disrupt the game balance of the virtual game it belongs to.

[0262] Optionally, in virtual shooting games where virtual shooting props are located, there are often many different types of virtual shooting prop main bodies, and each virtual shooting prop main body can be used as a virtual shooting prop to test or predict shooting results. In addition, virtual shooting games often have many different types of virtual shooting prop accessories, and virtual shooting prop main bodies and virtual shooting prop accessories can be flexibly combined to form new virtual shooting props. This means that testing or predicting shooting results requires not only testing all types of virtual shooting prop main bodies, but also testing new virtual shooting props obtained through any combination method, which dramatically increases the number of objects to be predicted or tested, and the pressure of prediction or testing also increases.

[0263] Optionally, in this embodiment, whether it is a virtual shooting prop body that is used alone as a virtual shooting prop, or a virtual shooting prop body and virtual shooting prop accessories that are used in combination as virtual shooting props, they can be regarded as prop data with the same or different shooting attribute characteristics, without distinguishing the type or combination method. Therefore, even if there are a large number of objects to be predicted or tested, there is no prediction or testing pressure in this embodiment.

[0264] Optionally, in this embodiment, the shooting attribute features may be, but are not limited to, attribute features in the shooting game scene that are highly correlated with the actual shooting results, such as the damage multiplier of a virtual bullet hitting the head of an enemy virtual object, the damage multiplier of a virtual bullet hitting the torso of an enemy virtual object, the firing interval (such as the time interval between two bullets firing in burst mode), the vertical recoil of the shot (base value, modification factor, maximum value), the horizontal recoil of the shot (base value, modification factor, maximum value), the time to put away the virtual shooting prop (the time required to raise the gun and put it down), and the time to raise the virtual shooting prop, etc.

[0265] Optionally, in this embodiment, the actual shooting result can be used, but is not limited to, to represent the combat result fed back during one or more rounds of shooting games using virtual shooting props, such as kill result, kill result, assist result, etc. Furthermore, predicting the shooting result of a virtual shooting prop can be understood, but is not limited to, predicting the combat result that the virtual shooting prop should feed back in each future round of shooting games.

[0266] Optionally, in this embodiment, the classification tree model may include, but is not limited to, a root node (non-leaf node), one or more leaf nodes, and one or more child nodes (non-leaf nodes). The root node can be understood as an input node, that is, classifying the shooting attribute features input to the classification tree model and outputting the classification result to the next node (which can be a child node or a leaf node). If the next node is a corresponding leaf node, a preset value of the corresponding leaf node is output, and the output result of the current classification tree model is determined based on this preset data. Optionally, in the classification tree model, each non-leaf node may have, but is not limited to, a preset classification condition, and this corresponding classification condition is used to classify the input features for output to the next node.

[0267] Optionally, in this embodiment, a set of classification tree models includes multiple classification tree models, and the target classification tree model selected from the set of classification tree models may include, but is not limited to, multiple classification tree models. In other words, the selection criteria in a set of classification tree models may include, but is not limited to, selecting all classification tree models that meet the preset conditions as the target classification tree model; furthermore, the target classification tree model selected from a set of classification tree models may include, but is not limited to, a single classification tree model. In other words, the selection criteria in a set of classification tree models may include, but is not limited to, selecting the classification tree model that best meets the preset conditions as the target classification tree model.

[0268] It should be noted that for ordinary prediction scenarios, model technology is undoubtedly the best way to ensure the accuracy and efficiency of prediction. However, the prediction scenarios of actual shooting results are different from ordinary prediction scenarios. The prediction scenarios of actual shooting results have high scenario complexity (there are many and complex factors affecting the actual shooting results, such as the rate of fire will change at different ranges, and the damage caused by the shooting will also change depending on the hit site). Therefore, how to apply model technology to the prediction scenarios of actual shooting results is still a difficult problem that related technologies have not yet overcome.

[0269] In this embodiment, to overcome this challenge, a classification tree model in model technology is proposed to predict actual shooting results. Furthermore, by setting shooting attribute features and corresponding classification conditions from the shooting attribute feature set on the non-leaf nodes of the classification tree model, shooting attribute features and corresponding classification conditions that closely match the shooting result prediction scenario are combined into the classification tree model, resulting in a set of classification tree models with a high degree of fit with the shooting result prediction scenario. The training process of the classification tree model is further improved by setting the average value of actual shooting results from the training sample subset on the leaf nodes. Finally, a target classification tree model is selected from the set of classification tree models for actual shooting result prediction, ensuring the accuracy of the shooting result prediction output by the model.

[0270] For specific implementation examples, please refer to the example shown in the training device of the shooting result prediction model described above, which will not be repeated here.

[0271] Through the embodiments provided in this application, a training sample set corresponding to a set of virtual shooting props is obtained. Each training sample in the training sample set includes a shooting attribute feature set of a virtual shooting prop from the set and the actual shooting result obtained using a virtual shooting prop. The shooting attribute feature set includes shooting attribute features of multiple dimensions. Based on the training sample set, a set of classification tree models is generated. The non-leaf nodes of the classification tree model are set with shooting attribute features from the shooting attribute feature set and corresponding classification conditions. The leaf nodes of the classification tree model are set with target values ​​of the actual shooting results in the training sample subset. The target values ​​include the average result of the actual shooting results in the training sample subset and the average residual of the difference between the actual shooting results and the average result in the training sample subset. The shooting attribute feature set of each virtual shooting prop in the training sample subset satisfies the classification conditions corresponding to the parent node of the leaf node. The training sample set includes a training sample subset. Based on the target values ​​set on the leaf nodes of the classification tree model and the actual shooting results in the training sample subset, a classification tree model is generated. A target classification tree model is selected. This model generates predicted shooting results for virtual shooting props based on shooting attribute features and corresponding classification conditions set on non-leaf nodes, and target values ​​set on leaf nodes. This model technology provides an efficient means of predicting actual shooting results. Furthermore, by combining the classification tree model with the shooting result prediction scenario with a high degree of fit, the accuracy of the actual shooting result prediction is ensured. This achieves the technical objective of improving both the efficiency and accuracy of actual shooting result prediction, thus providing a highly efficient and accurate means of predicting actual shooting results.

[0272] As an optional solution, generation unit 1404 includes:

[0273] The first selection module is used to select a first shooting attribute feature from the shooting attribute feature set according to the feature value distribution of each shooting attribute feature in the shooting attribute feature set in the training sample set.

[0274] The first generation module is used to generate a first classification tree model group with the node corresponding to the first shooting attribute feature as the root node. The leaf nodes in the first classification tree model group are set with the average result. A group of classification tree models includes the first classification tree model group.

[0275] The second selection module is used to select a second shooting attribute feature from the shooting attribute feature set based on the feature value distribution of shooting attribute features other than the first shooting attribute feature in the shooting attribute feature set in the training sample set.

[0276] The second generation module is used to generate a second classification tree model group with the node corresponding to the second shooting attribute feature as the root node. The leaf nodes in the second classification tree model group are set with residual average values. A set of classification tree models includes the second classification tree model group.

[0277] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0278] As an optional solution, selection unit 1406 includes:

[0279] The third selection module is used to select the first target classification tree model in the first classification tree model group based on the average result set on the leaf nodes in the first classification tree model group and the actual shooting results in the training sample subset. The first target classification tree model is used to generate the first predicted shooting result of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes in the first target classification tree model and the corresponding classification conditions, as well as the average result set on the leaf nodes.

[0280] The fourth selection module is used to select the second target classification tree model in the second classification tree model group based on the residual average value set on the leaf nodes in the second classification tree model group and the difference between the actual shooting results and the average result in the training sample subset. The second target classification tree model is used to generate the second predicted shooting result of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes in the second target classification tree model and the corresponding classification conditions, as well as the residual average value set on the leaf nodes.

[0281] The predicted shooting result of the target virtual shooting prop is the sum of the first predicted shooting result and the second predicted shooting result.

[0282] As an optional solution, the selected modules include:

[0283] The first determining submodule is used to determine the classification conditions corresponding to each shooting attribute feature in the shooting attribute feature set based on the feature value distribution of each shooting attribute feature in the shooting attribute feature set in the training sample set.

[0284] The first selection submodule is used to select a first shooting attribute feature from the shooting attribute feature set according to the classification conditions corresponding to each shooting attribute feature. The multiple training sample subsets are training sample subsets obtained from the training samples in the training sample set according to the classification conditions corresponding to the first shooting attribute feature. The feature value of the first shooting attribute feature in each of the multiple training sample subsets corresponds to a feature value cluster.

[0285] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0286] As an optional approach, the first generation module includes:

[0287] The first setting submodule is used to set the node corresponding to the first shooting attribute feature as the root node in the first classification tree model to be generated;

[0288] The first classification submodule is used to classify the training samples in the training sample set according to the classification conditions corresponding to the first shooting attribute features, and obtain multiple training sample subsets.

[0289] The second setting submodule is used to set the node corresponding to each training sample subset in multiple training sample subsets as a leaf node in the first classification tree model. The leaf node in the first classification tree model is a child node of the root node in the first classification tree model, and the target value set on the leaf node in the first classification tree model is the average value of the actual shooting results in the corresponding training sample subset.

[0290] The first determining submodule is used to determine whether to generate a second classification tree model based on the error between the target value set on the leaf node of the first classification tree model and the actual shooting result in the corresponding training sample subset.

[0291] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0292] As an optional approach, a second classification tree model group is generated using the node corresponding to the second shooting attribute feature as the root node, including:

[0293] The third setting submodule is used to set the node corresponding to the second shooting attribute feature as the root node in the third classification tree model to be generated;

[0294] The second classification submodule is used to classify the training samples in the training sample set according to the classification conditions corresponding to the second shooting attribute features, and obtain multiple training sample subsets.

[0295] The fourth setting submodule is used to set the node corresponding to each training sample subset in multiple training sample subsets as a leaf node in the third classification tree model. The leaf node in the third classification tree model is a child node of the root node in the third classification tree model. The target value set on the leaf node in the third classification tree model is the average value of the residual between the actual shooting result in the corresponding training sample subset and the average value of the actual shooting result in the training sample subset.

[0296] The second determining submodule is used to determine whether to generate a fourth classification tree model based on the error between the target value set on the leaf node of the third classification tree model and the residual value.

[0297] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0298] As an optional approach, a second classification tree model group is generated using the node corresponding to the second shooting attribute feature as the root node, including:

[0299] The third setting submodule is used to set the node corresponding to the second shooting attribute feature as the root node in the third classification tree model to be generated;

[0300] The second classification submodule is used to classify the training samples in the training sample set according to the classification conditions corresponding to the second shooting attribute features, and obtain multiple training sample subsets.

[0301] The fourth setting submodule is used to set the node corresponding to each training sample subset in multiple training sample subsets as a leaf node in the third classification tree model. The leaf node in the third classification tree model is a child node of the root node in the third classification tree model. The target value set on the leaf node in the third classification tree model is the average value of the residual between the actual shooting result in the corresponding training sample subset and the average value of the actual shooting result in the training sample subset.

[0302] The second determining submodule is used to determine whether to generate a fourth classification tree model based on the error between the target value set on the leaf node of the third classification tree model and the residual value.

[0303] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0304] As an optional approach, the sub-units are determined, including:

[0305] The sub-determination module is used to determine the first leaf node in the first classification tree model, wherein the error between the average value set on the first leaf node and the actual shooting result in the training sample subset corresponding to the first leaf node is greater than a third preset threshold.

[0306] The sub-selection module is used to select a second shooting attribute feature from the shooting attribute feature set based on the feature value distribution of shooting attribute features other than the first shooting attribute feature in the shooting attribute feature set in the training sample set. Here, multiple training sample subsets are training sample subsets obtained by training samples in the training sample set according to the classification conditions corresponding to the second shooting attribute feature. The feature value of the second shooting attribute feature in each of the multiple training sample subsets corresponds to a feature value cluster.

[0307] The sub-setting module is used to replace the first leaf node in the first classification tree model with the child node corresponding to the second shooting attribute feature, and set the node corresponding to each training sample subset in multiple training sample subsets as the leaf node of the child node corresponding to the second shooting attribute feature, so as to obtain the second classification tree model.

[0308] As an optional approach, the first generation module includes:

[0309] The execution submodule, used to perform the following steps when the node depth in the i-th classification tree model is less than the fourth preset threshold, is a natural number:

[0310] The third determining submodule is used to determine the generation of the (i+1)th classification tree model when the error between the average value set on the leaf node of the i-th classification tree model and the actual shooting result in the corresponding training sample subset is greater than the third preset threshold.

[0311] The third determination submodule is used to determine the first leaf node in the i-th classification tree model, wherein the error between the average value set on the first leaf node and the actual shooting result in the training sample subset corresponding to the first leaf node is greater than the third preset threshold.

[0312] The second selection submodule is used to select the (i+1)th shooting attribute feature from the shooting attribute feature set in the training sample set according to the feature value distribution of the shooting attribute features other than the first shooting attribute feature to the i-th shooting ray feature. Here, multiple training sample subsets are training sample subsets obtained by training samples in the training sample set according to the classification conditions corresponding to the (i+1)th shooting attribute feature. The feature value of the (i+1)th shooting attribute feature in each training sample subset corresponds to a feature value cluster.

[0313] The replacement submodule is used to replace the first leaf node in the i-th classification tree model with the i+1-th child node corresponding to the i+1-th shooting attribute feature, and set the node corresponding to each training sample subset in multiple training sample subsets as the leaf node of the i+1-th child node, thus obtaining the i+1-th classification tree model.

[0314] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0315] As an optional solution, it includes:

[0316] The acquisition submodule is used to obtain the error between the average value set on the leaf node of the (i+1)th child node and the actual shooting result in the training sample subset corresponding to the leaf node of the (i+1)th child node after obtaining the (i+1)th classification tree model.

[0317] The adjustment submodule is used to adjust the classification conditions corresponding to the shooting attribute feature set on the (i+1)th child node when there is a leaf node with an error greater than the third preset threshold in the leaf node of the (i+1)th child node after obtaining the (i+1)th classification tree model.

[0318] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0319] As an alternative solution,

[0320] The selection module includes: a third selection submodule, used to select N shooting attribute features from the set of shooting attribute features, where N is a natural number greater than 1;

[0321] The first generation module includes a generation submodule, which is used to generate a set of classification tree models corresponding to each of the N shooting attribute features in parallel, with the node corresponding to one of the shooting attribute features as the root node, resulting in a total of M sets of classification tree models. Here, a set of classification tree models includes M sets of classification tree models, where M is a natural number greater than or equal to N and greater than 1.

[0322] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0323] As an optional approach, submodules are generated, including:

[0324] The first generation subunit is used to generate, in parallel, a set of classification tree models corresponding to each of the N shooting attribute features, with the node corresponding to one of the shooting attribute features as the root node, resulting in a total of N sets of classification tree models; or,

[0325] The second generation subunit is used to generate a set of classification tree models corresponding to each of the N shooting attribute features in parallel, with the node corresponding to one of the shooting attribute features as the root node and the nodes corresponding to one or more of the other shooting attribute features as the child nodes of the root node, resulting in a total of M sets of classification tree models.

[0326] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0327] As an optional solution, generation unit 1404 includes:

[0328] The fifth selection module is used to select the current shooting attribute feature from the set of shooting attribute features;

[0329] The third generation module is used to generate the current classification tree model with the node corresponding to the current shooting attribute feature as the root node and the current classification condition corresponding to the current shooting attribute feature as the classification condition corresponding to the root node. The leaf nodes in the current classification tree model group are set with the current value, which includes the result average or the residual average.

[0330] The sixth selection module is used to select the next shooting attribute feature from the shooting attribute feature set as the current shooting attribute feature when the error corresponding to the current value has not reached the convergence condition, or to select the next classification condition corresponding to the current shooting attribute feature as the current classification condition, or to select the next shooting attribute feature from the shooting attribute feature set as the current shooting attribute feature and to select the classification condition corresponding to the next shooting attribute feature as the current classification condition.

[0331] The second determining module is used to determine a set of classification tree models, including the current classification tree model, when the error corresponding to the current value reaches the convergence condition.

[0332] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0333] According to another aspect of the present invention, a shooting result prediction device for implementing the above-described shooting result prediction method is also provided. For example... Figure 15 As shown, the device includes:

[0334] The second acquisition unit 1502 is used to acquire the target shooting attribute feature set of the target virtual shooting prop to be predicted;

[0335] The first determining unit 1504 is used to determine the target leaf node corresponding to the target virtual shooting prop in the leaf node of the target classification tree model according to the shooting attribute features set on the non-leaf nodes in the target classification tree model and the corresponding classification conditions, as well as the target shooting attribute feature set, wherein the target shooting attribute feature set satisfies the classification conditions set on the parent node of the target leaf node.

[0336] The third acquisition unit 1506 is used to acquire the target value set on the target leaf node, wherein the target value includes the average value of the actual shooting results and the average residual of the difference between the actual shooting results and the average value of the results.

[0337] The second determining unit 1508 is used to determine the predicted shooting result of the target virtual shooting prop based on the target value set on the target leaf node;

[0338] The target classification tree model is selected from a set of classification tree models based on the target values ​​set on the leaf nodes of a set of classification tree models and the actual shooting results in the corresponding training sample subset. The set of classification tree models is generated based on the training sample set. Each training sample in the training sample set includes the shooting attribute feature set of a virtual shooting prop in the virtual shooting prop set and the actual shooting result obtained by using a virtual shooting prop. The shooting attribute feature set includes shooting attribute features in multiple dimensions. The training sample set includes a training sample subset.

[0339] Optionally, in this embodiment, the shooting result prediction device can be applied, but is not limited to, in scenarios where the battle balance information of virtual shooting props is tested (used to measure whether the virtual shooting props disrupt the game balance of the virtual game in which the virtual shooting props are located due to excessively high / low prop strength). For example, a training sample set composed of multiple virtual shooting props as samples is first obtained, and each of the multiple virtual shooting props as samples carries historical usage data or is directly marked with one or more actual shooting results; then, the training sample set is used to train a classification tree model, and a set of classification tree models is obtained through training; further, to ensure the accuracy of the model output, a target classification tree model is selected from the set of classification tree models as the actual prediction model;

[0340] Based on this, the item data (such as item type, item model, item accessories, etc.) of the virtual shooting item to be tested is input into the target classification tree model, and the output of the target classification tree model is used as the predicted shooting result of the virtual shooting item to be tested. Then, based on the obtained predicted shooting result, the balance of the virtual shooting item to be predicted is judged (the judgment result can be understood as balance information, but is not limited to). Based on the judgment result, the item data of the virtual shooting item to be predicted is adjusted (or not adjusted) to ensure that the virtual shooting item to be predicted will not disrupt the game balance of the virtual game it belongs to.

[0341] Optionally, in virtual shooting games where virtual shooting props are located, there are often many different types of virtual shooting prop main bodies, and each virtual shooting prop main body can be used as a virtual shooting prop to test or predict shooting results. In addition, virtual shooting games often have many different types of virtual shooting prop accessories, and virtual shooting prop main bodies and virtual shooting prop accessories can be flexibly combined to form new virtual shooting props. This means that testing or predicting shooting results requires not only testing all types of virtual shooting prop main bodies, but also testing new virtual shooting props obtained through any combination method, which dramatically increases the number of objects to be predicted or tested, and the pressure of prediction or testing also increases.

[0342] Optionally, in this embodiment, whether it is a virtual shooting prop body that is used alone as a virtual shooting prop, or a virtual shooting prop body and virtual shooting prop accessories that are used in combination as virtual shooting props, they can be regarded as prop data with the same or different shooting attribute characteristics, without distinguishing the type or combination method. Therefore, even if there are a large number of objects to be predicted or tested, there is no prediction or testing pressure in this embodiment.

[0343] Optionally, in this embodiment, the shooting attribute features may be, but are not limited to, attribute features in the shooting game scene that are highly correlated with the actual shooting results, such as the damage multiplier of a virtual bullet hitting the head of an enemy virtual object, the damage multiplier of a virtual bullet hitting the torso of an enemy virtual object, the firing interval (such as the time interval between two bullets firing in burst mode), the vertical recoil of the shot (base value, modification factor, maximum value), the horizontal recoil of the shot (base value, modification factor, maximum value), the time to put away the virtual shooting prop (the time required to raise the gun and put it down), and the time to raise the virtual shooting prop, etc.

[0344] Optionally, in this embodiment, the actual shooting result can be used, but is not limited to, to represent the combat result fed back during one or more rounds of shooting games using virtual shooting props, such as kill result, kill result, assist result, etc. Furthermore, predicting the shooting result of a virtual shooting prop can be understood, but is not limited to, predicting the combat result that the virtual shooting prop should feed back in each future round of shooting games.

[0345] Optionally, in this embodiment, the classification tree model may include, but is not limited to, a root node (non-leaf node), one or more leaf nodes, and one or more child nodes (non-leaf nodes). The root node can be understood as an input node, that is, classifying the shooting attribute features input to the classification tree model and outputting the classification result to the next node (which can be a child node or a leaf node). If the next node is a corresponding leaf node, a preset value of the corresponding leaf node is output, and the output result of the current classification tree model is determined based on this preset data. Optionally, in the classification tree model, each non-leaf node may have, but is not limited to, a preset classification condition, and this corresponding classification condition is used to classify the input features for output to the next node.

[0346] Optionally, in this embodiment, a set of classification tree models includes multiple classification tree models, and the target classification tree model selected from the set of classification tree models may include, but is not limited to, multiple classification tree models. In other words, the selection criteria in a set of classification tree models may include, but is not limited to, selecting all classification tree models that meet the preset conditions as the target classification tree model; furthermore, the target classification tree model selected from a set of classification tree models may include, but is not limited to, a single classification tree model. In other words, the selection criteria in a set of classification tree models may include, but is not limited to, selecting the classification tree model that best meets the preset conditions as the target classification tree model.

[0347] It should be noted that for ordinary prediction scenarios, model technology is undoubtedly the best way to ensure the accuracy and efficiency of prediction. However, the prediction scenarios of actual shooting results are different from ordinary prediction scenarios. The prediction scenarios of actual shooting results have high scenario complexity (there are many and complex factors affecting the actual shooting results, such as the rate of fire will change at different ranges, and the damage caused by the shooting will also change depending on the hit site). Therefore, how to apply model technology to the prediction scenarios of actual shooting results is still a difficult problem that related technologies have not yet overcome.

[0348] In this embodiment, to overcome this challenge, a classification tree model in model technology is proposed to predict actual shooting results. Furthermore, by setting shooting attribute features and corresponding classification conditions from the shooting attribute feature set on the non-leaf nodes of the classification tree model, shooting attribute features and corresponding classification conditions that closely match the shooting result prediction scenario are combined into the classification tree model, resulting in a set of classification tree models with a high degree of fit with the shooting result prediction scenario. The training process of the classification tree model is further improved by setting the average value of actual shooting results from the training sample subset on the leaf nodes. Finally, a target classification tree model is selected from the set of classification tree models for actual shooting result prediction, ensuring the accuracy of the shooting result prediction output by the model.

[0349] For specific implementation examples, please refer to the example shown in the training device of the shooting result prediction model described above, which will not be repeated here.

[0350] The embodiments provided in this application obtain a set of target shooting attribute features for the virtual shooting prop to be predicted; according to the shooting attribute features set on the non-leaf nodes of the target classification tree model and the corresponding classification conditions, and the set of target shooting attribute features, the target leaf node corresponding to the virtual shooting prop is determined in the leaf nodes of the target classification tree model, wherein the set of target shooting attribute features satisfies the classification conditions set on the parent node of the target leaf node; the target value set on the target leaf node is obtained, wherein the target value includes the average value of the actual shooting result and the average residual of the difference between the actual shooting result and the average value of the result; based on the target value set on the target leaf node, the predicted shooting result of the virtual shooting prop is determined. This provides an efficient means of predicting actual shooting results through model technology, and also ensures the accuracy of the actual shooting result prediction by combining the classification tree model with the shooting result prediction scenario with high fit. Thus, it achieves the technical objective of improving the prediction efficiency of actual shooting results while ensuring the prediction accuracy of actual shooting results, thereby realizing the technical effect of providing an efficient and accurate means of predicting actual shooting results.

[0351] As an alternative solution,

[0352] The third acquisition unit 1506 includes: a second acquisition module, used to acquire the average result and the average residual set on the target leaf node in the P classification tree models when the target classification tree model includes P classification tree models, where P is a natural number greater than 1;

[0353] The second determining unit 1508 includes: a first determining module, used to determine the predicted shooting result of the target virtual shooting prop as equal to the sum of the average result and the average residual.

[0354] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0355] As an optional solution, the first determining unit 1504 includes:

[0356] The search module is used to search for nodes in the target classification tree model that satisfy the classification conditions of the target shooting attribute feature set, starting from the root node, until the target leaf node is found.

[0357] For specific implementation examples, please refer to the example shown in the training method of the shooting result prediction model above, which will not be repeated here.

[0358] According to another aspect of the present invention, an electronic device for implementing the training method of the above-described shooting result prediction model is also provided, such as... Figure 16 As shown, the electronic device includes a memory 1602 and a processor 1604. The memory 1602 stores a computer program, and the processor 1604 is configured to execute the steps of any of the above method embodiments via the computer program.

[0359] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0360] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0361] S1, obtain the training sample set corresponding to the virtual shooting prop set, wherein each training sample in the training sample set includes the shooting attribute feature set of a virtual shooting prop in the virtual shooting prop set and the actual shooting result obtained by using a virtual shooting prop. The shooting attribute feature set includes shooting attribute features of multiple dimensions.

[0362] S2, Based on the training sample set, generate a set of classification tree models. In the classification tree model, the non-leaf nodes are set with shooting attribute features from the shooting attribute feature set and the corresponding classification conditions. The leaf nodes are set with the target values ​​of the actual shooting results in the training sample subset. The target values ​​include the average value of the actual shooting results in the training sample subset and the average residual value of the difference between the actual shooting results and the average value in the training sample subset. The shooting attribute feature set of each virtual shooting prop in the training sample subset satisfies the classification conditions corresponding to the parent node of the leaf node. The training sample set includes the training sample subset.

[0363] S3, based on the target values ​​set on the leaf nodes of the classification tree model and the actual shooting results in the training sample subset, select a target classification tree model from a set of classification tree models. The target classification tree model is used to generate predicted shooting results for the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes of the target classification tree model, the corresponding classification conditions, and the target values ​​set on the leaf nodes; or,

[0364] S1, Obtain the target shooting attribute feature set of the virtual shooting prop to be predicted;

[0365] S2, according to the shooting attribute features and corresponding classification conditions set on the non-leaf nodes in the target classification tree model, and the target shooting attribute feature set, determine the target leaf node corresponding to the target virtual shooting prop in the leaf nodes of the target classification tree model, wherein the target shooting attribute feature set satisfies the classification conditions set on the parent node of the target leaf node;

[0366] S3, obtain the target value set on the target leaf node, where the target value includes the average value of the actual shooting results and the average residual of the difference between the actual shooting results and the average value of the results;

[0367] S4, based on the target value set on the target leaf node, determine the predicted shooting result of the target virtual shooting prop;

[0368] The target classification tree model is selected from a set of classification tree models based on the target values ​​set on the leaf nodes of a set of classification tree models and the actual shooting results in the corresponding training sample subset. The set of classification tree models is generated based on the training sample set. Each training sample in the training sample set includes the shooting attribute feature set of a virtual shooting prop in the virtual shooting prop set and the actual shooting result obtained by using a virtual shooting prop. The shooting attribute feature set includes shooting attribute features in multiple dimensions. The training sample set includes a training sample subset.

[0369] Alternatively, as those skilled in the art will understand, Figure 16 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones (such as Android phones, iOS phones, etc.), tablets, PDAs, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 16 This does not limit the structure of the aforementioned electronic devices. For example, the electronic device may also include components that are more... Figure 16 The more or fewer components shown (such as network interfaces, etc.), or having the same Figure 16 The different configurations shown.

[0370] The memory 1602 can be used to store software programs and modules, such as the program instructions / modules corresponding to the training method and apparatus for the shooting result prediction model in this embodiment of the invention. The processor 1604 executes various functional applications and data processing by running the software programs and modules stored in the memory 1602, thereby realizing the aforementioned training method for the shooting result prediction model. The memory 1602 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1602 may further include memory remotely located relative to the processor 1604, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Specifically, the memory 1602 may be used, but is not limited to, to store training sample sets, target classification tree models, and information such as predicted shooting results. As an example, such as Figure 16 As shown, the memory 1602 may include, but is not limited to, the first acquisition unit 1402, generation unit 1404, and selection unit 1406 of the training device for the shooting result prediction model; the memory 1602 may also include, but is not limited to, the second acquisition unit 1502, first determination unit 1504, third acquisition unit 1506, and second determination unit 1508 of the shooting result prediction device. Furthermore, other module units of the training device for the shooting result prediction model may also be included, but will not be described further in this example.

[0371] Optionally, the transmission device 1606 described above is used to receive or send data via a network. Specific examples of the network described above may include wired networks and wireless networks. In one example, the transmission device 1606 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 1606 is a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0372] In addition, the aforementioned electronic device also includes: a display 1608 for displaying information such as the aforementioned training sample set, target classification tree model, and predicted shooting results; and a connection bus 1610 for connecting the various module components in the aforementioned electronic device.

[0373] In other embodiments, the aforementioned terminal device or server can be a node in a distributed system, wherein the distributed system can be a blockchain system, which is a distributed system formed by connecting multiple nodes through network communication. The nodes can form a peer-to-peer (P2P) network, and any form of computing device, such as a server, terminal, or other electronic device, can become a node in the blockchain system by joining this peer-to-peer network.

[0374] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned method for predicting shooting results and the aforementioned method for training a shooting result prediction model, wherein the computer program is configured to execute the steps of any of the aforementioned method embodiments during runtime.

[0375] Optionally, in this embodiment, the computer-readable storage medium described above may be configured to store a computer program for performing the following steps:

[0376] S1, obtain the training sample set corresponding to the virtual shooting prop set, wherein each training sample in the training sample set includes the shooting attribute feature set of a virtual shooting prop in the virtual shooting prop set and the actual shooting result obtained by using a virtual shooting prop. The shooting attribute feature set includes shooting attribute features of multiple dimensions.

[0377] S2, Based on the training sample set, generate a set of classification tree models. In the classification tree model, the non-leaf nodes are set with shooting attribute features from the shooting attribute feature set and the corresponding classification conditions. The leaf nodes are set with the target values ​​of the actual shooting results in the training sample subset. The target values ​​include the average value of the actual shooting results in the training sample subset and the average residual value of the difference between the actual shooting results and the average value in the training sample subset. The shooting attribute feature set of each virtual shooting prop in the training sample subset satisfies the classification conditions corresponding to the parent node of the leaf node. The training sample set includes the training sample subset.

[0378] S3, based on the target values ​​set on the leaf nodes of the classification tree model and the actual shooting results in the training sample subset, select a target classification tree model from a set of classification tree models. The target classification tree model is used to generate predicted shooting results for the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes of the target classification tree model, the corresponding classification conditions, and the target values ​​set on the leaf nodes; or,

[0379] S1, Obtain the target shooting attribute feature set of the virtual shooting prop to be predicted;

[0380] S2, according to the shooting attribute features and corresponding classification conditions set on the non-leaf nodes in the target classification tree model, and the target shooting attribute feature set, determine the target leaf node corresponding to the target virtual shooting prop in the leaf nodes of the target classification tree model, wherein the target shooting attribute feature set satisfies the classification conditions set on the parent node of the target leaf node;

[0381] S3, obtain the target value set on the target leaf node, where the target value includes the average value of the actual shooting results and the average residual of the difference between the actual shooting results and the average value of the results;

[0382] S4, based on the target value set on the target leaf node, determine the predicted shooting result of the target virtual shooting prop;

[0383] The target classification tree model is selected from a set of classification tree models based on the target values ​​set on the leaf nodes of a set of classification tree models and the actual shooting results in the corresponding training sample subset. The set of classification tree models is generated based on the training sample set. Each training sample in the training sample set includes the shooting attribute feature set of a virtual shooting prop in the virtual shooting prop set and the actual shooting result obtained by using a virtual shooting prop. The shooting attribute feature set includes shooting attribute features in multiple dimensions. The training sample set includes a training sample subset.

[0384] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0385] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0386] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more computer devices (which may be personal computers, servers, or network devices, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention.

[0387] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0388] In the several embodiments provided in this application, it should be understood that the disclosed client can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0389] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0390] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0391] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A training method for a shooting result prediction model, characterized in that, include: Obtain a training sample set corresponding to the set of virtual shooting props, wherein each training sample in the training sample set includes a shooting attribute feature set of a virtual shooting prop in the set of virtual shooting props and the actual shooting result obtained by using the virtual shooting prop, and the shooting attribute feature set includes shooting attribute features of multiple dimensions; Based on the training sample set, a set of classification tree models is generated. The non-leaf nodes of the classification tree model are set with shooting attribute features from the shooting attribute feature set and corresponding classification conditions. The leaf nodes of the classification tree model are set with target values ​​of actual shooting results from the training sample subset. These target values ​​include the average result of the actual shooting results from the training sample subset and the average residual of the difference between the actual shooting results from the training sample subset and the average result. The shooting attribute feature set of each virtual shooting prop in the training sample subset satisfies the classification conditions corresponding to the parent node of the leaf node. The training sample set includes the training sample subset. Based on the target values ​​set on the leaf nodes of the classification tree model and the actual shooting results in the training sample subset, a target classification tree model is selected from the set of classification tree models. The target classification tree model is used to generate the predicted shooting results of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes of the target classification tree model and the corresponding classification conditions, as well as the target values ​​set on the leaf nodes.

2. The method according to claim 1, characterized in that, The step of generating a set of classification tree models based on the training sample set includes: Based on the feature value distribution of each shooting attribute feature in the shooting attribute feature set in the training sample set, a first shooting attribute feature is selected from the shooting attribute feature set. Using the node corresponding to the first shooting attribute feature as the root node, a first classification tree model group is generated, wherein the leaf nodes in the first classification tree model group are set with the average value of the result, and the group of classification tree models includes the first classification tree model group. Based on the feature value distribution of the shooting attribute features other than the first shooting attribute feature in the shooting attribute feature set in the training sample set, a second shooting attribute feature is selected in the shooting attribute feature set; Using the node corresponding to the second shooting attribute feature as the root node, a second classification tree model group is generated. The leaf nodes in the second classification tree model group are set with the residual average value. The group of classification tree models includes the second classification tree model group.

3. The method according to claim 2, characterized in that, The step of selecting a target classification tree model from the set of classification tree models based on the target values ​​set on the leaf nodes of the classification tree model and the actual shooting results in the training sample subset includes: Based on the average result set on the leaf nodes in the first classification tree model group and the actual shooting results in the training sample subset, a first target classification tree model is selected in the first classification tree model group. The first target classification tree model is used to generate the first predicted shooting result of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes and the corresponding classification conditions, as well as the average result set on the leaf nodes in the first target classification tree model. Based on the residual average value set on the leaf nodes in the second classification tree model group and the difference between the actual shooting results and the average value of the results in the training sample subset, a second target classification tree model is selected in the second classification tree model group. The second target classification tree model is used to generate the second predicted shooting result of the target virtual shooting prop based on the shooting attribute features set on the non-leaf nodes and the corresponding classification conditions, as well as the residual average value set on the leaf nodes in the second target classification tree model. The predicted shooting result of the target virtual shooting prop is the sum of the first predicted shooting result and the second predicted shooting result.

4. The method according to claim 2, characterized in that, The step of selecting a first shooting attribute feature from the shooting attribute feature set based on the feature value distribution of each shooting attribute feature in the shooting attribute feature set in the training sample set includes: Based on the feature value distribution of each shooting attribute feature in the shooting attribute feature set in the training sample set, determine the classification conditions corresponding to each shooting attribute feature in the shooting attribute feature set; Based on the classification conditions corresponding to each shooting attribute feature, the first shooting attribute feature is selected from the set of shooting attribute features. The plurality of training sample subsets are training sample subsets obtained from the training samples in the training sample set according to the classification conditions corresponding to the first shooting attribute feature. The feature value of the first shooting attribute feature in each of the plurality of training sample subsets corresponds to a feature value cluster.

5. The method according to claim 2, characterized in that, The step of generating a first classification tree model group by using the node corresponding to the first shooting attribute feature as the root node includes: Set the node corresponding to the first shooting attribute feature as the root node in the first classification tree model to be generated; The training samples in the training sample set are classified according to the classification conditions corresponding to the first shooting attribute feature to obtain multiple training sample subsets; Each node corresponding to a training sample subset in the plurality of training sample subsets is set as a leaf node in the first classification tree model, wherein the leaf node in the first classification tree model is a child node of the root node in the first classification tree model, and the target value set on the leaf node in the first classification tree model is the average value of the actual shooting results in the corresponding training sample subset. Based on the error between the target value set on the leaf node of the first classification tree model and the actual shooting result in the corresponding training sample subset, determine whether to generate a second classification tree model.

6. The method according to claim 2, characterized in that, The step of generating a second classification tree model group by using the node corresponding to the second shooting attribute feature as the root node includes: Set the node corresponding to the second shooting attribute feature as the root node in the third classification tree model to be generated; The training samples in the training sample set are classified according to the classification conditions corresponding to the second shooting attribute features to obtain multiple training sample subsets; Each node corresponding to a training sample subset in the plurality of training sample subsets is set as a leaf node in the third classification tree model, wherein the leaf node in the third classification tree model is a child node of the root node in the third classification tree model, and the target value set on the leaf node in the third classification tree model is the average value of the residual values ​​between the actual shooting results in the corresponding training sample subset and the average value of the actual shooting results in the training sample subset. Based on the error between the target value set on the leaf node of the third classification tree model and the residual value, determine whether to generate a fourth classification tree model.

7. The method according to claim 5, characterized in that, The step of determining whether to generate a second classification tree model based on the error between the target value set on the leaf node of the first classification tree model and the actual shooting result in the corresponding training sample subset includes: When the error between the average value set on the leaf node of the first classification tree model and the actual shooting result in the corresponding training sample subset is greater than a third preset threshold, the second classification tree model is determined to be generated, and the second classification tree model is generated based on the first classification tree model.

8. The method according to claim 7, characterized in that, The step of generating the second classification tree model based on the first classification tree model includes: In the first classification tree model, a first leaf node is determined, wherein the error between the average value set on the first leaf node and the actual shooting result in the training sample subset corresponding to the first leaf node is greater than the third preset threshold. Based on the feature value distribution of the shooting attribute features other than the first shooting attribute feature in the shooting attribute feature set in the training sample set, a second shooting attribute feature is selected in the shooting attribute feature set. The multiple training sample subsets are training sample subsets obtained by training samples in the training sample set according to the classification conditions corresponding to the second shooting attribute feature. The feature value of the second shooting attribute feature in each of the multiple training sample subsets corresponds to a feature value cluster. The first leaf node in the first classification tree model is replaced with the child node corresponding to the second shooting attribute feature, and the node corresponding to each training sample subset in the plurality of training sample subsets is set as the leaf node of the child node corresponding to the second shooting attribute feature, thus obtaining the second classification tree model.

9. The method according to claim 7, characterized in that, The step of generating the set of classification tree models using the node corresponding to the first shooting attribute feature as the root node includes: If the node depth in the i-th classification tree model is less than the fourth preset threshold, perform the following steps, where i is a natural number: When the error between the average value set on the leaf node of the i-th classification tree model and the actual shooting result in the corresponding training sample subset is greater than the third preset threshold, the (i+1)-th classification tree model is determined to be generated. In the i-th classification tree model, a first leaf node is determined, wherein the error between the average value set on the first leaf node and the actual shooting result in the training sample subset corresponding to the first leaf node is greater than the third preset threshold. Based on the feature value distribution of the shooting attribute features in the shooting attribute feature set in the training sample set, excluding the first shooting attribute feature to the i-th shooting ray feature, the (i+1)-th shooting attribute feature is selected in the shooting attribute feature set. Multiple training sample subsets are training sample subsets obtained from the training samples in the training sample set according to the classification conditions corresponding to the (i+1)-th shooting attribute feature. The feature value of the (i+1)-th shooting attribute feature in each of the multiple training sample subsets corresponds to a feature value cluster. The first leaf node in the i-th classification tree model is replaced with the (i+1)-th child node corresponding to the (i+1)-th shooting attribute feature, and the node corresponding to each training sample subset in the plurality of training sample subsets is set as the leaf node of the (i+1)-th child node to obtain the (i+1)-th classification tree model.

10. The method according to claim 9, characterized in that, After obtaining the (i+1)th classification tree model, the following steps are included: In the (i+1)th classification tree model, the error between the average value set on the leaf node of the (i+1)th child node and the actual shooting result in the training sample subset corresponding to the leaf node of the (i+1)th child node is obtained; When there is a leaf node in the (i+1)th child node whose error is greater than the third preset threshold, the classification conditions corresponding to the (i+1)th shooting attribute feature set on the (i+1)th child node are adjusted.

11. The method according to claim 2, characterized in that, The step of selecting the first shooting attribute feature from the shooting attribute feature set includes: selecting N shooting attribute features from the shooting attribute feature set, where N is a natural number greater than 1; The step of generating the set of classification tree models with the node corresponding to the first shooting attribute feature as the root node includes: generating a set of classification tree models corresponding to each shooting attribute feature in parallel with the node corresponding to one of the N shooting attribute features as the root node, resulting in a total of M sets of classification tree models, wherein the set of classification tree models includes the M sets of classification tree models, and M is a natural number greater than or equal to N and greater than 1.

12. The method according to claim 11, characterized in that, The process involves using the node corresponding to one of the N shooting attribute features as the root node to generate a set of classification tree models corresponding to each shooting attribute feature in parallel, resulting in a total of M sets of classification tree models, including: Using the node corresponding to one of the N shooting attribute features as the root node, a set of classification tree models corresponding to each shooting attribute feature is generated in parallel, resulting in a total of N sets of classification tree models; or, Each of the N shooting attribute features is used as a root node, and the node corresponding to one or more of the other shooting attribute features is used as a child node of the root node. In parallel, a set of classification tree models corresponding to each shooting attribute feature is generated, resulting in a total of M sets of classification tree models.

13. The method according to any one of claims 1 to 12, characterized in that, The step of generating a set of classification tree models based on the training sample set includes: Select the current shooting attribute feature from the set of shooting attribute features; Using the node corresponding to the current shooting attribute feature as the root node, and the current classification condition corresponding to the current shooting attribute feature as the classification condition corresponding to the root node, a current classification tree model is generated. The leaf nodes in the current classification tree model group are set with current values, including the average result or the average residual. If the error corresponding to the current value has not reached the convergence condition, the next shooting attribute feature is selected from the shooting attribute feature set as the current shooting attribute feature, or the next classification condition corresponding to the current shooting attribute feature is selected from the shooting attribute feature set as the current classification condition, or the next shooting attribute feature is selected from the shooting attribute feature set as the current shooting attribute feature and the classification condition corresponding to the next shooting attribute feature is selected as the current classification condition. If the error corresponding to the current value reaches the convergence condition, then the set of classification tree models is determined to include the current classification tree model.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program, when executed, performs the method described in any one of claims 1 to 13.

15. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 13 through the computer program.

Images