An AI-based multi-mode tactical shooting simulation training method

Abstract

The application discloses a kind of multi-mode tactical shooting simulation training methods based on AI, including real-time acquisition of trainee shooting data, through hit rate, threat suppression time, ammunition economy and risk exposure index calculation comprehensive shooting efficiency score, build update tactical proficiency matrix;Based on the score, the ability state of trainee is determined using Bayesian probability inference, and information entropy is calculated to select training mode;In repair mode, through reverse reinforcement learning, generate demonstration type AI opponent scene, in test mode, build habitual strategy model and optimize disturbance parameter to generate suppression scene;Using matrix evolution trend, the core parameters are optimized by evolution algorithm, to realize the self iteration of system.The present application solves the fundamental problem that the prior art cannot be self-optimized, relies on manual review and is low in efficiency, and improves the tactical thinking and adaptability of the trainee.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tactical shooting simulation training, and particularly relates to a multi-mode tactical shooting simulation training method based on AI. BACKGROUND

[0002] With the rapid development of computer graphics, physics engines and artificial intelligence technology, virtual simulation training systems have been widely used in high-risk professional fields such as military, police and emergency response. These systems aim to provide a repeatable, low-cost and absolutely safe training platform for trainees by building a high-fidelity virtual environment. In the current tactical shooting simulation training technology, artificial intelligence mainly plays the core role of driving non-player character behavior and generating dynamic confrontation scenarios. By applying behavior trees, finite state machines and even preliminary reinforcement learning technologies, existing systems have achieved a certain degree of simulation of battlefield environment, significantly improving the immersion and confrontation of shooting simulation, and providing an effective way to improve the tactical decision-making and shooting ability of trainees.

[0003] However, despite the progress made by existing technologies, there are still some limitations. First, in the dimension of shooting training evaluation, existing systems mostly use coarse-grained evaluation indicators such as hit rate and completion time, which makes it difficult to conduct in-depth attribution diagnosis of trainees' ability weaknesses and cannot accurately answer the core question of "why performance is poor". Second, in the adaptability of shooting training content, the difficulty adjustment mechanism of most systems is relatively rigid, usually limited to simple linear logic such as increasing the number of enemies or improving their basic attributes, lacking the ability to generate highly relevant tactical scenarios for individual specific weaknesses, resulting in low efficiency of shooting training. Third, the behavior patterns of existing AI opponents, even if complex, are often fixed or predictable, but trainees can easily form path dependence and tactical solidification after familiarization, which is contrary to the uncertainty of real battlefield and is not conducive to the cultivation of tactical thinking and adaptability. Fourth, existing shooting training systems are static, their evaluation standards and training logic are set based on trainers, lacking an optimization mechanism that can feedback according to long-term training results, resulting in the overall effectiveness of the shooting training system being unable to achieve self-optimization and low efficiency of review.

[0004] CN120071698A discloses a military police VR simulation training system and device, which mainly uses virtual reality technology to realistically simulate combat environments in various modes (such as target range training mode and task execution mode, etc.), so that the training personnel feel as if they are in a real battlefield, thereby effectively improving the realism and immersion of training. Although the scheme mentions training AI to simulate enemy characters, the training function is a "black box" and the processing process cannot be known. In addition, the training process does not have intelligent identification of continuous optimization and trainee ability short board, and cannot achieve intelligent training, but is reviewed by the trainer, which brings great challenges to the energy consumption of the trainer and also makes it difficult to ensure the accuracy of the training review. SUMMARY

[0005] This section is intended to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification of the present application in order to avoid obscuring the purpose of this section, the abstract and the title, and such simplifications or omissions cannot be used to limit the scope of the present application.

[0006] In view of the above existing problems, the present application is proposed. Therefore, the present application provides a multi-mode tactical shooting simulation training method based on AI to solve the problems proposed in the background art.

[0007] To solve the above technical problems, the present application provides the following technical solutions: a multi-mode tactical shooting simulation training method based on AI, comprising:

[0008] In the simulation training, the shooting data of the trainee is obtained in real time, a comprehensive shooting efficiency score is obtained by considering the shooting evaluation index of the trainee, and a tactical proficiency matrix recording the score is constructed and updated;

[0009] Based on the obtained comprehensive shooting efficiency score, a probability reasoning method is used to determine the ability state of the trainee in a specific tactical situation, and according to the ability state, a training mode is selected from the preset repair mode and test mode;

[0010] According to the selected training mode, the corresponding tactical situation is generated, if the repair mode is selected, the first type of tactical situation of the AI opponent is generated, and if the test mode is selected, the second type of tactical situation of the AI opponent is generated;

[0011] The evolution trend of the tactical proficiency matrix over time is used to optimize the core parameters in the above steps by an evolution algorithm.

[0012] As a preferred scheme of the AI-based multi-mode tactical shooting simulation training method, the shooting data of the trainee obtained in real time at least includes target type, target state, shooting distance, and environmental conditions.

[0013] As a preferred scheme of the AI-based multi-mode tactical shooting simulation training method, the shooting examination index of the trainee at least includes hit rate, threat suppression time, ammunition economy, and risk exposure.

[0014] As a preferred scheme of the AI-based multi-mode tactical shooting simulation training method, the quantification process of the risk exposure includes:

[0015] During the shooting period, the instantaneous threat value of a target weapon basic threat value, distance decay effect from the target unit, continuous pressure effect from the target aiming, and trainee's own shooting shelter reduction coefficient are integrated in time.

[0016] As a preferred scheme of the AI-based multi-mode tactical shooting simulation training method, the determination of the ability state of the trainee in a specific tactical scenario by using a probabilistic reasoning method includes:

[0017] The posterior probability of the ability state belonging to a preset short board state, proficiency state, and mastery state is calculated by using a Bayesian inference method.

[0018] As a preferred scheme of the AI-based multi-mode tactical shooting simulation training method, it further includes:

[0019] Before selecting the training mode according to the posterior probability, the information entropy of the posterior probability distribution is calculated, and when the information entropy is higher than a preset information entropy threshold, a standardized tactical scenario is generated.

[0020] As a preferred scheme of the AI-based multi-mode tactical shooting simulation training method, the process of updating the shooting efficiency score uses an exponential moving average algorithm, and the smoothing factor of the exponential moving average algorithm is dynamically adjusted according to the trend of the ability state of the trainee in a specific tactical scenario.

[0021] As a preferred scheme of the AI-based multi-mode tactical shooting simulation training method, the behavior pattern generated by the AI opponent is generated by modeling the ability state, including:

[0022] By defining an idealized expert behavior trajectory corresponding to the ability state, an reward function of the trajectory is inferred in reverse, and an intelligent agent is trained based on the reward function.

[0023] As a preferred scheme of the AI-based multi-mode tactical shooting simulation training method according to the application, wherein, if the test mode is selected, a second type of tactical scene of the AI opponent is generated, and the method further comprises:

[0024] A habitual strategy model of the trainee is constructed, and a scene disturbance parameter is found by a gradient optimization method, and a tactical element in the second type of tactical scene is modified according to the scene disturbance parameter.

[0025] As a preferred scheme of the AI-based multi-mode tactical shooting simulation training method according to the application, wherein, an optimization target of the evolutionary algorithm is to maximize a fitness function, and evaluation of the fitness function includes at least an overall mean value improvement speed and an overall variance reduction speed of the tactical proficiency matrix.

[0026] Compared with the prior art, the application has the following beneficial effects:

[0027] 1. The application quantifies and tracks the comprehensive ability of the trainee in different tactical scenes by constructing and continuously updating a tactical proficiency matrix that integrates multi-dimensional indicators such as hit rate and risk exposure, thereby accurately understanding the ability short board of the trainee and solving the problem of coarse evaluation dimension in the prior art and the inability to perform deep attribution.

[0028] 2. By constructing a dual-mode AI scene generation mechanism of repair mode and test mode, scenes that compensate for the ability short board of the trainee are generated using inverse reinforcement learning, and test scenes that break tactical inertia are generated, thereby solving the problem of AI logic solidification, insufficient adaptability, and difficulty in effectively improving the tactical thinking and response ability of the trainee in the prior art.

[0029] 3. The long-term evolution trend of the tactical proficiency matrix is used as the fitness function of the evolutionary algorithm, so that the training method itself can continuously evolve to adapt to changing training needs and data patterns, thereby solving the fundamental problem of the prior art that cannot be self-optimized, relies on manual review, and is low in efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings. Among them:

[0031] Figure 1 The overall flowchart of the AI-based multi-mode tactical shooting simulation training method according to an embodiment of the application;

[0032] Figure 2 The evolutionary algorithm execution flow chart of the AI-based multi-mode tactical shooting simulation training method is described for an embodiment of the present application. DETAILED DESCRIPTION

[0033] In order to make the above objectives, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application are described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0034] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be practiced without the specific details that are set forth in the description, and it will be apparent to those skilled in the art that the present application can be practiced with other than the specific details set forth in the description. Therefore, the present application is not limited to the specific embodiments disclosed in the description.

[0035] Secondly, the "one embodiment" or "embodiment" referred to herein can include specific features, structures or characteristics that can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is separate or alternative to other embodiments.

[0036] The present application is described in detail in conjunction with the schematic diagram, and in the detailed description of the embodiments of the present application, the cross-sectional view of the device structure is locally enlarged without the general proportion for the convenience of description, and the schematic diagram is only an example, which should not limit the scope of protection of the present application herein. In addition, three-dimensional spatial dimensions including length, width and depth should be included in actual manufacture.

[0037] Meanwhile, in the description of the present application, it should be noted that the terms "up, down, inner and outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first, second or third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0038] Unless otherwise defined, the terms "mounting, connecting, associating" in the present application should be interpreted broadly, for example: it can be fixed connection, detachable connection or integral connection; it can also be mechanical connection, electrical connection or direct connection, it can also be indirectly connected through intermediate medium, or it can be internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0039] Embodiment 1

[0040] Reference Figure 1 and Figure 2 As the first embodiment of the present application, the embodiment provides an AI-based multi-mode tactical shooting simulation training method, comprising:

[0041] S1, in the simulation training, the shooting data of the trainee is acquired in real time, a comprehensive shooting efficiency score is obtained by considering the shooting evaluation index of the trainee, and a tactical proficiency matrix recording the score is constructed and updated.

[0042] Further, in the process of shooting simulation training of the trainee, high-frequency (for example, at a frequency of 60 times per second or higher) data stream, i.e. original shooting data, is captured in real time by interacting with the underlying engine of the shooting simulation training system.

[0043] Further, the original shooting data includes but is not limited to target type, target state, shooting distance and environmental conditions.

[0044] Specifically, the target type refers to the specific target type when the trainee shoots, for example, stationary target, moving target, simulated enemy combatants, light armored vehicles or specific high-value threat objects, etc.

[0045] Specifically, the target state refers to the behavior state of the shooting target when it faces the trainee shooting, for example, whether the shooting target is stationary, moving at a constant speed, moving at a variable speed, seeking cover, or is aiming / shooting at the trainee, etc.

[0046] It should be noted that the target type and the target state together constitute the definition of "tactical problem", which is not only the unified concept in the traditional method, i.e. "enemy / enemy army", but also includes identifying the specific unit of "enemy / enemy army" (for example, "moving infantry", "high-speed impact light vehicle", etc.).

[0047] Specifically, the shooting distance is calculated and recorded in real time by calculating the dynamic Euclidean distance between the trainee and each shooting target.

[0048] It should be noted that the shooting distance is easily affected by environmental conditions.

[0049] Specifically, the environmental conditions refer to the macro and micro environmental parameters at the time of shooting, such as the light conditions of day / night, weather conditions of sunny / rainy / snowy, and reduced visibility caused by smoke or dust, etc.

[0050] It should be noted that the shooting distance and the environmental conditions jointly define the constraint conditions of the "tactical problem", for example, shooting at a distance of 200 meters at night using a weapon with low-light aiming to test the skill set of the trainee under different combat conditions, by taking the shooting distance and the environmental conditions as the key variables, rather than irrelevant backgrounds, a multi-class target scenario is constructed based on the original shooting data.

[0051] In addition, the traditional shooting simulation training system usually records data in the form of isolated event logs (such as "10:01:03-Trainee A shoots", "10:06:39-Target Unit B is eliminated"), which loses the causality and concurrency relationship between events and cannot form a complete "shooting" concept, for example, it cannot be determined whether the trainee's shooting is the cause of the target unit's elimination, nor can it be known whether other target units are aiming at the trainee at the same time during the shooting. The fragmentation of such data is the fundamental reason for the coarse granularity of the evaluation dimension. In order to solve the above problems, the method introduces an event stream processing engine based on a finite state machine (Finite State Machine, FSM) into the underlying engine of the shooting simulation training system. The engine is mainly used to reconstruct the context of shooting training, specifically, the identification and segmentation of shooting state.

[0052] Specifically, the state machine defines at least three core states: standby, shooting, and disengagement. By monitoring key trigger data in real time (such as the trainee entering the target perception range, the trainee shooting through the weapon, the target attacking the trainee, etc.), the state machine dynamically identifies the starting point and ending point of a complete "shooting" in the data stream, which divides the continuous data stream into multiple discrete "tactical slices" with clear start and end boundaries. Then, for each "tactical slice", the method will forcibly associate all related data within its time window, including but not limited to: determining the main participants of this shooting (the trainee and one or more specific targets), capturing the continuous displacement trajectory of all participants during the shooting, recording each shooting event and its corresponding ballistic result, associating the environmental light and weather parameters at the time of shooting, and synchronously recording the aiming state of all target units against the trainee.

[0053] It should be noted that by introducing the finite state machine-based event stream processing engine, the original, unordered data points can be reconstructed into a series of structured, information-complete "tactical slices", each of which can be represented as a detailed shooting report containing the shooting assessment indicators of the trainee. Through the "tactical slice", meaningful and process-based indicator calculation can be performed.

[0054] Further, the required data directly extracted from the "tactical slice" includes but is not limited to hit rate, threat suppression time, ammunition economy (ammunition consumption) and risk exposure.

[0055] In addition, since the prior art cannot determine whether the trainee is aiming at one or more targets at a certain moment, the risk assessment in the traditional sense can only be based on the result statistics after the trainee is hit by the target, which has serious lag and contingency.

[0056] Specifically, based on the above problems, especially the synchronous recording of the target unit aiming state, within the time span of each "tactical slice", the risk exposure indicator no longer focuses on the result "whether the trainee is hit" but on the process "to what extent the trainee is exposed to danger".

[0057] Further, by time-integrating an instantaneous threat value (i.e., the target unit is first locked to the threat and the threat is removed) that comprehensively considers the target weapon basic threat value (e.g., rifles, artillery, etc., with different basic threat coefficients), the distance decay effect (i.e., the closer the distance, the greater the threat), the continuous pressure effect of being aimed at by the target (i.e., the longer the time of being aimed at, the higher the potential risk) and the trainee's own shooting shelter reduction coefficient (i.e., different shelter materials provide damage reduction coefficients), a cumulative risk total is calculated.

[0058] Specifically, the risk total can be expressed by the formula:

[0059] ,

[0060] wherein, is the basic threat value of the target weapon, which is a preset constant (e.g., 10 for a rifle and 50 for artillery), is the distance decay function, which is an inverse function , is the real-time distance, is the distance decay coefficient, ranging from 0 to 1, is the continuous pressure function, which is a linear growth function , is the duration of the trainee being aimed at, is a pressure growth coefficient, ranging from [0, 1], is a damage reduction coefficient of the shelter where the trainee is currently located, which is a constant in the interval [0, 1] (for example, 0.6 for an earth wall and 0.9 for a concrete shelter).

[0061] It should be noted that through the cumulative total risk, the traditional method is changed from "trainee damage results" to "trainee risk management behavior", so that the actual survival ability of the trainee in the process of tactical shooting and shooting with cover can be truly reflected, and the hysteresis and contingency are avoided.

[0062] Furthermore, a tactical proficiency matrix is constructed to stably track the real trend of the trainee's ability.

[0063] Specifically, the constructed tactical proficiency matrix is a multi-dimensional matrix, the dimensions of which are defined by the dimensions of the aforementioned original shooting data. Each element in the matrix stores the (weighted and normalized) comprehensive shooting efficiency score of the trainee under the combination of the ith target scenario and the jth environmental condition, which makes the ability of the trainee no longer a specific score, but a clear and visual ability spectrum, and the "short board" and "advantage" of the ability are clear at a glance on the matrix.

[0064] Specifically, the comprehensive shooting efficiency score can be expressed by the formula:

[0065] ,

[0066] wherein, represents the hit rate, represents the threat suppression time, represents the ammunition economy, represents the risk exposure, , , , and respectively represent the weight coefficients of the above four indexes, and the sum of the weight coefficients is 1.

[0067] ​In addition, the comprehensive shooting efficiency score calculated after each shooting simulation training is filled in or updated in the corresponding element position of the tactical proficiency matrix, and an exponential moving average algorithm is adopted, that is, the new comprehensive shooting efficiency score is not directly overwritten on the old comprehensive shooting efficiency score, but the new comprehensive shooting efficiency score and the historical comprehensive shooting efficiency score in the element position are weighted and averaged according to a smoothing factor, so that each value in the tactical proficiency matrix can stably reflect the recent comprehensive ability level of the trainee in the corresponding tactical situation, thereby constituting an accurate, dynamic and multi-dimensional personal ability change trend of the trainee.

[0068] S2, based on the obtained comprehensive shooting efficiency score, a probability reasoning method is used to determine the ability state of the trainee in the specific tactical situation, and according to the ability state, one of the training modes is selected from the preset repair mode and test mode.

[0069] It should be noted that since the traditional shooting simulation training system usually compares the comprehensive shooting efficiency score of the trainee with a fixed threshold when judging the score, the fundamental defect of this method is that it uses a "deterministic" way of thinking, so it cannot handle the inherent "uncertainty" problem in the data. For example, a trainee obtains a comprehensive shooting efficiency score of 65 points in a certain tactical situation. In fact, this score itself is ambiguous: the trainee may be a student whose ability is stable at the 65-point level, or an excellent student whose average level is 80 points but performs abnormally in this tactical situation, or a student whose level is 50 points but performs abnormally in this tactical situation. Therefore, the difficulty adjustment for the trainee is often blind, and may even have negative effects (for example, considering that the excellent student who performs abnormally is actually lacking in personal ability).

[0070] Specifically, in order to solve the above problems, the present application introduces a probability reasoning method to determine the true ability state of the trainee, that is, under the condition that the trainee's score this time is "XX", the probability of the trainee's true ability state belonging to the short board, proficiency or mastery of the several preset states is respectively how much.

[0071] Specifically, the ability of the trainee in a certain tactical situation (i.e. an element position in the tactical proficiency matrix) is divided into a state space composed of at least three discrete states: short-plate state, proficient state and master state. After obtaining a comprehensive shooting efficiency score, the method does not directly use the score, but uses it as "evidence" to update the posterior probability of the system on the ability state of the trainee in the corresponding tactical situation. This updating process follows the Bayes formula, i.e. the posterior probability of the ability of the trainee belonging to each of the above three preset states is calculated under the condition of obtaining the current comprehensive shooting efficiency score. For example, for the score of 65 points, assuming P(state = short-plate | score = 65) = 0.1, P(state = proficient | score = 65) = 0.6, P(state = master | score = 65) = 0.3, then through the posterior probability distribution, the following conclusion can be obtained: although the score is not high, the real ability state of the trainee is most likely to be in the proficient state rather than the short-plate state.

[0072] Specifically, the Bayes formula is represented as:

[0073]

[0074] wherein, is the prior probability, which is obtained in the following manner: when the shooting simulation training system accepts a new trainee, the prior probability is first adopted in the manner of non-informative prior, for example, {P(state = short-plate) = 1 / 3, P(state = proficient) = 1 / 3, P(state = master) = 1 / 3}. With the running of the system, the posterior probability calculated by the system in the last round is directly used as the prior probability for the calculation of the system in the next round, is the likelihood probability, which is obtained by statistical analysis of the scaled historical training data, i.e. fitting a probability density function, is the ability state of the trainee.

[0075] Preferably, the probability density function can be realized by using a Gaussian Mixture Model (GMM). For each ability state (short-plate, proficient and master) of the trainee, the system will collect all historical comprehensive shooting efficiency scores in the state, and use the Expectation-Maximization (EM) algorithm to fit a GMM for the scores. A GMM model can be composed of multiple Gaussian distributions. When a new score is obtained, the system only needs to substitute the score into the GMM model corresponding to the state, so as to obtain an accurate likelihood probability value.

[0076] ​It should be noted that the GMM was chosen instead of a single Gaussian distribution because it can better capture the multimodal distribution characteristics that may exist under specific ability states (e.g., low peak scores caused by occasional mistakes by trainees in a skilled state), thus making the calculation of likelihood probability more accurate.

[0077] Furthermore, even with the introduction of probabilistic reasoning, if the system faces an ambiguous probability distribution, such as P(state = weakness) = 0.48 and P(state = proficiency) = 0.52, the confidence level of the decision is very low regardless of which state is chosen. Existing methods are helpless in this regard and can only "bite the bullet" and make a potentially wrong choice. Therefore, they lack the ability to self-reflect and actively verify.

[0078] Specifically, in order to solve the above problems and before making a state selection, a self-checking mechanism is introduced. The core of this mechanism is to calculate the information entropy of the current posterior probability distribution.

[0079] It should be noted that, in the context of this invention, information entropy refers to the precise quantification of the "degree of uncertainty in drawing conclusions" of the system itself. An extremely high probability distribution (such as P(state = weakness) = 0.95) corresponds to an extremely low information entropy, indicating that the system highly accepts the conclusion that "the trainee has a weakness." Conversely, a stable probability distribution (such as the above P(state = weakness) = 0.48, P(state = proficiency) = 0.52) corresponds to an extremely high information entropy, indicating that the system itself is in a state of "cognitive confusion."

[0080] Furthermore, by pre-setting an information entropy threshold, when the information entropy of the calculated posterior probability distribution exceeds this threshold, the system will pause the selection of the regular training mode. This prevents the system from blindly engaging in a low-confidence game between "repair" and "test," and instead proactively takes an action that maximizes information gain: generating a standardized tactical scenario for the trainee (i.e., fixed environmental conditions, target type, target state, and shooting distance, with no adversarial intervention). This scenario is equivalent to a calibration test, the purpose of which is not training, but to obtain more discriminative performance data so that the information entropy can be significantly reduced in the next round of Bayesian inference, thereby allowing the system to draw more accurate conclusions.

[0081] Specifically, information entropy threshold The calculation method is as follows:

[0082] ,

[0083] in, The maximum possible information entropy (i.e., the entropy under uniform distribution) (where D is the number of capability states), and It is an adjustable hyperparameter, and its range is within... between.

[0084] Further, only when the information entropy is lower than the preset information entropy threshold, that is, the system has a high enough confidence in the conclusion drawn by itself, enters the decision stage, that is, the system selects the mode according to the state with the maximum probability value in the posterior probability distribution: if it is the short board state, the repair mode is selected, and if it is the skilled state or the proficient state, the test mode is selected.

[0085] It should be noted that after the mode selection, an instruction is transmitted at this time, which is used to guide the size of the smoothing factor in the S1 step, if it is the repair mode, it means that the ability of the trainee is in the promotion stage, a smaller smoothing factor is used when updating the tactical proficiency matrix, so that the tactical proficiency matrix is more sensitive to the recent comprehensive shooting efficiency score of the trainee, so as to reflect the progress of the trainee's ability, if it is the test mode, it means that the ability of the trainee tends to be stable, and the improvement space is small, a larger smoothing factor is used when updating the tactical proficiency matrix, so that the tactical proficiency matrix can effectively filter out the comprehensive shooting efficiency score fluctuation caused by accidental failure, so as to reflect the current true level of the trainee.

[0086] Specifically, the update of the smoothing factor can be expressed as:

[0087] ,

[0088] wherein, is the growth rate of the posterior probability of the trainee's ability state, if the probability is rapidly rising, it is a larger positive number, if it tends to be stable, it is close to 0, in the present application, is expressed as a mapping function, which is used to map the growth rate result to the interval (0, 1).

[0089] S3, generating a corresponding tactical scene according to the selected training mode, if the repair mode is selected, the first type of tactical scene of the AI opponent is generated, and if the test mode is selected, the second type of tactical scene of the AI opponent is generated.

[0090] It should be noted that when the traditional system identifies the short board of the trainee (for example, the score is low on “medium distance attack moving target”), the usual method is to let the trainee repeatedly train the same type of scene, which is essentially “trial and error”, which only provides the trainee with “problems”, but does not provide “problem solving ideas”, the trainee may accidentally succeed after many failures, but what he learns may be a suboptimal solution, or even an incorrect solution that is only effective under the current parameters, so the process is inefficient.

[0091] Specifically, to solve the above problems, the repair mode of the application, the core idea is not to let the trainee himself explore, but to "demonstrate" through the behavior of the AI opponent, that is, by generating an AI opponent, the behavior pattern of the AI opponent itself is the embodiment of the "standard solution" to the current trainee's short board.

[0092] Further, for each short board state of the trainee corresponding to the tactical scene, the method collects idealized expert behavior trajectories in two ways:

[0093] The first kind, invite experienced tactical experts to demonstrate operation in the target scene of the shooting simulation training system for many times, and record the complete state-action sequence data stream, that is, state-action pairs .

[0094] Specifically, the state includes but is not limited to: the three-dimensional coordinates of the expert, the posture (standing / squatting / lying), the current life value, the remaining ammunition, and the types, positions and states of all target units in its field of view.

[0095] Specifically, the action includes but is not limited to: the distance of self-moving, the screen coordinate change of the shooting point, the shooting instruction, and the ammunition replacement instruction.

[0096] The second kind, an optimal strategy agent in the scene is trained in advance through large-scale offline reinforcement learning, and its behavior trajectory is recorded.

[0097] Specifically, the idealized expert behavior trajectory is not a simple moving route, but a high-dimensional time series data, in which each frame contains: {environment state, trainee's own state, observed target state}→{target moving distance, target aiming, shooting decision, trainee's posture switching, shooting equipment usage}.

[0098] It should be noted that the idealized expert behavior trajectory represents a typical example in the target scene.

[0099] In addition, it is extremely difficult and fragile to directly build a behavior tree of an AI opponent that can perfectly execute the idealized expert behavior trajectory described above, therefore, the application adopts the reverse reinforcement learning method.

[0100] Further, by taking the expert behavior trajectory as input, a reward function that can explain the trajectory is inferred in reverse.

[0101] It should be noted that the reward function is essentially a mathematical expression of expert behavior, for example, in the target scenario, "stabilizing the sight for more than 0.5 seconds before shooting" will get a high reward, "exposing more than 1 second in an unshielded area" will be punished, "using short bursts instead of full-automatic sweeping" will get a reward, etc.

[0102] Further, when the reward function is inferred, a new agent is trained based on the reward function through a standard reinforcement learning algorithm (such as PPO), which maximizes the reward function through continuous trial and error in the shooting simulation training environment. The AI opponent formed by the final convergence strategy will naturally exhibit highly consistent tactical behavior with the expert. In the repair mode, the trainee faces the AI opponent, and the shooting distance, state, and tactical thinking of the trainee are exemplary. Under these conditions, the trainee can intuitively feel and learn efficient shooting tactics by competing with the AI opponent, thereby repairing the short board of their ability.

[0103] It should be noted that for trainees with proficient or expert ability, traditional systems usually challenge them by linearly increasing the difficulty (more, faster, stronger targets). This method can only test the reaction speed and operation proficiency of the trainee, but cannot effectively challenge their tactical thinking. A proficient / expert trainee often forms a fixed and effective tactical strategy, and the traditional system cannot identify this "tactical solidification" and cannot create a scenario that can specifically counter the strategy, resulting in high-level shooting simulation training becoming a monotonous "score brushing", which is not conducive to cultivating tactical flexibility and adaptability.

[0104] Specifically, to solve the above problems, the core goal of the test mode of the present application is to identify and break the trainee's tactical comfort zone, aiming to generate a "nemesis" scenario that has the maximum counteracting effect on the trainee's tactical strategy, and this process follows the logic of adversarial attack.

[0105] Further, using the historical "tactical slice" data of the trainee in similar scenarios accumulated in step S1, a habitual strategy model that can mimic and predict the behavior of the trainee is constructed through behavior cloning or sequence modeling (such as LSTM) technology. This model can be understood as a specific implementation of the trainee's tactical strategy. When given a target scenario, the model can output the tactical strategy that the trainee is most likely to take.

[0106] Further, the key tactical elements in the target scenario (e.g. the initial position of the target, the reinforcement route of the enemy, the distribution of available cover for the self, etc.) are parameterized to form a scenario parameter space, and then a gradient optimization process is initiated, which aims to find a scenario perturbation parameter (i.e. a fine-tuning of the scenario parameters) based on the scenario parameter space, so that in the perturbed scenario, the expected performance (e.g. the expected overall shooting efficiency score) of the habitual strategy model of the trainee behavior is the worst.

[0107] Specifically, the gradient optimization process can be characterized by a loss function, which aims to maximize the failure probability of the trainee strategy model in the perturbed scenario (i.e. the expected performance of the habitual strategy model of the trainee behavior is the worst), and the loss function can be formally represented as:

[0108] ,

[0109] wherein, is the habitual strategy model of the trainee behavior, is the scenario perturbation parameter defined target scenario, is the predicted probability of the trainee achieving the task goal in the scenario.

[0110] Specifically, the gradient optimization adopts the gradient ascent method, which obtains:

[0111] ,

[0112] wherein, denotes the learning rate of the gradient ascent, is the gradient of the loss function with respect to the scenario perturbation parameter, is the current gradient, denotes the next gradient.

[0113] It should be noted that this process is mathematically equivalent to generating an "adversarial sample" for the strategy model of the trainee, and the optimization process will fine-tune the scenario parameters along the gradient direction that can most quickly reduce the expected performance of the model, and finally the optimization process will converge to an optimal scenario perturbation parameter.

[0114] Further, the scenario disturbance parameters found by the above optimization process are applied to a target scenario, thereby accurately finding a scenario configuration that can suppress the habitual strategy of the trainee, and generating a final "natural enemy" scenario as an AI opponent. For example, if the habitual strategy model of the trainee shows that it always prefers to occupy the left / right side high ground, the optimization process will generate a scenario in which crossfire is preset on the high ground, so that when the trainee rushes to the left / right side high ground, he will find that his "tactical strategy optimal solution" has become a "death trap".

[0115] It should be noted that in this way, the test mode can force the trainee to abandon his solidified tactical strategy, re-examine the battlefield environment, and immediately come up with a new tactical solution, greatly exercising the trainee's on-the-spot response ability and tactical flexibility, and achieving high-level shooting simulation training from "consolidating proficiency" to "improving tactical cognitive upper limit".

[0116] S4, using the evolution trend of the tactical proficiency matrix over time, optimizing the core parameters in the above steps through an evolutionary algorithm.

[0117] It should be noted that, in the foregoing steps, the weights in the comprehensive shooting efficiency score, the posterior probability of Bayesian inference, the threshold of information entropy, and the reward function of reinforcement learning are set by experts or developers when the shooting simulation training system is shipped, and the underlying training logic and evaluation criteria are fixed and unchanged regardless of how many trainees the system has trained and how much data it has accumulated. Therefore, when new weapons and equipment, new tactical theories, or trainee groups with changed ability characteristics are introduced, the entire system cannot adjust itself to adapt to the new environment, resulting in a gradual decline in training effectiveness, and long (i.e., the trainer reviews the trainee) and inefficient manual parameter adjustment.

[0118] Specifically, to solve the above problem, a meta-learning layer is introduced, which is implemented through an evolutionary algorithm framework simulating "natural selection".

[0119] Further, the weights in the above comprehensive shooting efficiency score, the posterior probability of Bayesian inference, the threshold of information entropy, and the reward function of reinforcement learning are abstracted as a core parameter gene group (Core Parameter Genome, CPG), and each CPG represents a unique end-to-end training methodology.

[0120] Specifically, the CPG includes at least: an evaluation gene, i.e., the weighting coefficient of the comprehensive shooting efficiency score, a diagnosis strategy gene, i.e., the information entropy threshold and the posterior probability distribution of the trainee's ability state in Bayesian inference, a scenario generation gene, i.e., the basis function of the reward function in inverse reinforcement learning and the learning rate in the habitual strategy model of the trainee's behavior.

[0121] Further, a fitness function is designed to evaluate the effectiveness of a CPG (i.e. a set of training methodology) after its application.

[0122] Specifically, the fitness function includes the overall mean improvement rate and the overall variance reduction rate, which can be expressed by the following formula:

[0123]

[0124] wherein, is the fitness function, is the learning efficiency, which represents the rate of change of the overall mean of the tactical proficiency matrix within a training period (e.g. the past 1000 training sessions) after the application of the CPG, is the learning quality, which represents the rate of change of the overall variance of the tactical proficiency matrix within the same training period, is the breakthrough reward, which represents the number of times all trainees exceed their historical highest difficulty level in the test mode, are the weights of the corresponding items, and the initial values are 0.5, 0.3, and 0.2, respectively.

[0125] Further, the overall mean improvement rate is calculated by the linear regression slope of the time series of the mean of the tactical proficiency matrix of all trainees within an evaluation period T (e.g. a week or 1000 training sessions), which represents the efficiency of the trainees' learning after the application of the set of training methodology (CPG).

[0126] Further, the overall variance reduction rate is calculated by the linear regression slope of the time series of the variance of the tactical proficiency matrix of all trainees within an evaluation period T, which represents the quality of the trainees' learning after the application of the set of training methodology (CPG).

[0127] It should be noted that the higher the values of the overall mean improvement rate and the overall variance reduction rate, the faster the learning efficiency and the higher the learning quality of the trainees.

[0128] In addition, due to the high-dimensional and non-linear complex parameter space of the CPG, traditional optimization methods such as gradient descent are prone to local optimization. However, by using the classical evolutionary algorithm through its group search and random exploration mechanism, the global optimal solution can be found more effectively, i.e. finding the optimal core parameter gene group that maximizes the fitness function, for reference Figure 2 The specific process is as follows:

[0129] ​​​S401, population initialization: generate N random CPG individuals, representing N distinct training methodologies.

[0130] S402, evaluate the generated CPG individuals: deploy N distinct training methodologies in parallel, let each training methodology run independently S1-S3, when the running is completed, calculate the fitness function of each CPG according to the evolution data of all collected tactical proficiency matrices, and obtain the corresponding fitness score.

[0131] S403, selection and crossover: select the top M CPG individuals with the highest fitness scores as "parents", and perform gene recombination on two selected "parent" CPGs to generate "child" CPG.

[0132] S404, mutation: apply random disturbance (minimum value) to the generated child CPG.

[0133] S405, iteration and convergence: replace the old population with the newly generated children, and return to S402 for continuous iteration.

[0134] It should be noted that after a plurality of iteration processes, the entire population gradually converges to the area with higher fitness scores, and finally outputs the individual with the highest fitness score as the optimal core parameter genome. The genome is used as the global default configuration of the shooting simulation training system when receiving new weapons, new tactical theories or facing the ability characteristics of the trainee group, so as to continuously evolve the system.

[0135] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages, such as object-oriented programming language Java and interpreted scripting language JavaScript.

[0136] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The flowchart and / or block diagram in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to the present application. In this regard, each block in the flowchart and / or block diagrams can represent a module, segment, or portion of code, which comprises one or more executable Figure 1 instructions for implementing the specified logical functions. It should also be noted that each block of the flowchart and / or block diagrams and that a flowchart and / or block diagram can represent a combination of

[0137] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 functionality specified in the flowchart block or blocks. The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1

[0138] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 Figure 1

[0139] While the preferred embodiments of the application have been described, additional variations and modifications can be employed by those skilled in the art. Therefore, the appended claims are intended to cover all such variations and modifications as falling within the scope of the application.

[0140] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.​​​

Claims

1. An AI-based multi-mode tactical shooting simulation training method, characterized in that, Comprise: In the simulation training, the shooting data of the trainee is acquired in real time, a comprehensive shooting efficiency score is obtained by considering the shooting evaluation index of the trainee, and a tactical proficiency matrix recording the score is constructed and updated; The shooting evaluation index of the trainee includes hit rate, threat suppression time, ammunition economy and risk exposure; The quantification process of risk exposure includes time integration of an instantaneous threat value which is a comprehensive of the target weapon basic threat value, distance decay effect from the target unit, continuous pressure effect from the target aiming and trainee's own shooting shelter reduction coefficient during the shooting period; Based on the obtained comprehensive shooting efficiency score, the ability state of the trainee in a specific tactical scenario is determined by using a probabilistic reasoning method, and a training mode is selected from the pre-set repair mode and test mode according to the ability state; According to the selected training mode, the corresponding tactical scenario is generated, if the repair mode is selected, the first type of tactical scenario of AI opponent is generated, if the test mode is selected, the second type of tactical scenario of AI opponent is generated; The behavior mode of the first type of tactical scenario AI opponent is generated by modeling the ability state, including defining an idealized expert behavior trajectory corresponding to the ability state, inversely inferring a reward function of the trajectory, and training an intelligent agent based on the reward function; If the test mode is selected, the second type of tactical scenario of AI opponent is generated, including constructing the habitual strategy model of the trainee, and finding a scenario disturbance parameter by gradient optimization method, and modifying the tactical elements in the second type of tactical scenario according to the scenario disturbance parameter; The evolution trend of the tactical proficiency matrix over time is used to optimize the core parameters in the above steps by evolutionary algorithm; The optimization target of the evolutionary algorithm is to maximize a fitness function, and the evaluation of the fitness function includes the overall mean improvement speed and the overall variance reduction speed of the tactical proficiency matrix. 2.The AI-based multi-mode tactical shooting simulation training method of claim 1, wherein, The real-time acquisition of the shooting data of the trainee includes target type, target state, shooting distance and environmental conditions. 3.The AI-based multi-mode tactical shooting simulation training method of claim 1, wherein, The ability state of the trainee in a specific tactical scenario is determined by using a probabilistic reasoning method, including: The posterior probability of the ability state belonging to the pre-set short board state, proficiency state and mastery state is calculated by Bayesian inference method. 4.The AI-based multi-mode tactical shooting simulation training method of claim 3, wherein, It also includes calculating the information entropy of the posterior probability distribution before selecting the training mode according to the posterior probability, and generating a standardized tactical scenario when the information entropy is higher than the pre-set information entropy threshold. 5.The AI-based multi-mode tactical shooting simulation training method according to claim 1 or 3, characterized in that, The process of updating the shooting efficiency score adopts exponential moving average algorithm, and the smoothing factor of the exponential moving average algorithm is dynamically adjusted according to the change trend of the ability state of the trainee in a specific tactical scenario.

Images