Game index and intention reasoning-based adversarial game intention prediction method and system

Abstract

The invention discloses an adversarial game intention prediction method and system based on game indexes and intention reasoning, and relates to the technical field of adversarial game decision making, and the method comprises the steps: receiving multi-source heterogeneous information and dynamic situation data of a game scene, the multi-source heterogeneous information comprising visual modal information, non-visual time sequence modal information and knowledge base modal information; the dynamic situation data of the game scene and a pre-constructed parameterized game index function library are subjected to relational mapping, a task completion evaluation result is obtained, and the pre-constructed parameterized game index function library comprises a plurality of task evaluation index functions associated with the specific confrontation game scene; the multi-source heterogeneous information is input into the pre-established game intention reasoning model, the intention recognition result is output and obtained, the task completion evaluation result is associated with the intention recognition result, prediction and evaluation of the confrontation game scene are achieved, and the recognition accuracy of the confrontation party cluster game intention is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of adversarial game decision-making technology, specifically to an adversarial game intent prediction method and system based on game indicators and intent reasoning. Background Technology

[0002] With the development of artificial intelligence, sensor networks, and multi-agent collaborative technologies, multi-agent systems are increasingly being used in dynamic and uncertain adversarial scenarios (such as production line adversarial scenarios in intelligent manufacturing and cluster games in highly adversarial intelligent games). However, current multi-agent situational analysis and collaborative decision-making technologies for adversarial games still have significant limitations.

[0003] First, in terms of game evaluation, most adversarial game systems currently use only one or a few fixed evaluation metrics (such as win rate and loss rate) to measure the adversarial effectiveness of each task or cluster. Such a single or fixed metric is difficult to adapt to the needs of different task types, different cluster structures, and different adversarial environments, resulting in simplistic evaluation results that are difficult to guide agents in selecting the optimal strategy.

[0004] Secondly, in terms of adversary intent prediction and collaborative behavior reasoning, existing methods often resort to flattening all modal information and uniformly inputting it into a traditional recurrent network (such as LSTM / GRU) when facing high-speed maneuvering adversary clusters, rapidly changing visual frame information, constantly changing speed and agent type composition, and the fusion of multi-dimensional probabilistic information in the situational knowledge base. This "flattening + single network" approach often ignores the spatiotemporal structural differences between modalities, the temporal correlation between information, and the spatial evolution characteristics, thus lacking the ability to fuse multi-source heterogeneous information and model the spatiotemporal evolution behavior of adversary clusters.

[0005] Furthermore, in terms of real-time dynamic decision-making and collaborative strategy optimization, existing systems struggle to simultaneously complete the closed-loop process of situational awareness, intent reasoning, game evaluation, and task strategy generation. Specifically, this manifests as: significant delays in responding to intelligence input, decreased prediction accuracy, lag in generating collaborative instructions, and reduced overall collaboration, thus hindering the efficient execution of adversarial game tasks by multiple agents under conditions of incomplete information.

[0006] Therefore, in complex, uncertain, and heterogeneous adversarial scenarios, existing technologies are inaccurate in predicting the opponent's intentions and suffer from low collaborative efficiency. Summary of the Invention

[0007] To address the shortcomings mentioned in the background art, the present invention aims to provide a method and system for predicting adversarial game intent based on game indicators and intent reasoning.

[0008] Firstly, the objective of this invention can be achieved through the following technical solution: a method for predicting intentions in adversarial games based on game indicators and intention reasoning, the method comprising the following steps: Receive multi-source heterogeneous information and dynamic situational data of the game scenario, wherein the multi-source heterogeneous information includes visual modal information, non-visual temporal modal information and knowledge base modal information; Among them, visual modal information is a continuous frame image sequence that reflects the changes in the size and distribution of the adversary cluster; non-visual temporal modal information is temporal data formed by changes in the adversary's direction of travel, speed, and agent type composition; and knowledge base modal information is prior knowledge and intelligence. The dynamic situational data of the game scenario is mapped to a pre-built parameterized game indicator function library to obtain the task completion evaluation result. The pre-built parameterized game indicator function library contains multiple task evaluation indicator functions associated with specific adversarial game scenarios. By inputting multi-source heterogeneous information into a pre-established game intent reasoning model, the intention recognition result is output. The task completion evaluation result is correlated with the intention recognition result, thus realizing the prediction and evaluation of adversarial game scenarios.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the continuous frame image sequence reflects the spatiotemporal evolution characteristics of the cluster, the frame rate is M frames / second, and the knowledge base modal information is obtained from the situational analysis knowledge base and converted into a probability vector of length k. .

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the parameterized game index function library is constructed based on expert knowledge; The multiple task evaluation index functions associated with specific adversarial game scenarios consist of multiple parameters, including but not limited to: winning time, winning rate, and local task success rate.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the pre-established game intent reasoning model includes a visual branch module, a non-visual branch module, a knowledge base branch module, and a hierarchical fusion module.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the visual branch module employs a three-dimensional convolutional neural network (3D CNN) to extract spatiotemporal features from visual modal information to obtain the feature vector output by the visual branch; The non-visual branch module uses a gated recurrent unit (GRU) to extract long-term features from non-visual temporal modal information and obtain the feature vector output by the non-visual branch. The knowledge base branch module is used to process prior knowledge base modal information in the form of probability vectors to obtain the feature vector output by the knowledge base branch. The hierarchical fusion module includes a spatiotemporal attention mechanism, which performs weighted fusion on the feature vectors of the visual branch output, the non-visual branch output, and the knowledge base branch output to obtain fused features. The fused features are then input into a fully connected layer to output the game intent recognition result of the adversary cluster.

[0013] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the convolution kernel of the 3D CNN of the visual branch module has a × b spatial dimension and c frames in temporal dimension; the number of GRU hidden units of the non-visual branch module is h; the spatiotemporal attention mechanism of the hierarchical fusion module includes: a temporal dimension attention submodule and a spatial dimension attention submodule, used to assign weights to temporal evolution features and spatial distribution features respectively.

[0014] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the game intent identification result of the adversary cluster is a set of intent categories {I1, I2, …} and the probability value corresponding to each intent category output by the fully connected layer of the hierarchical fusion module.

[0015] Secondly, in order to achieve the above objectives, this invention discloses an adversarial game intent prediction system based on game indicators and intent reasoning, comprising: The data receiving module is used to receive multi-source heterogeneous information and dynamic situational data of the game scenario. The multi-source heterogeneous information includes visual modal information, non-visual temporal modal information and knowledge base modal information. Among them, visual modal information is a continuous frame image sequence that reflects the changes in the size and distribution of the adversary cluster; non-visual temporal modal information is temporal data formed by changes in the adversary's direction of travel, speed, and agent type composition; and knowledge base modal information is prior knowledge and intelligence. The relation mapping module is used to map the dynamic situation data of the game scenario to a pre-built parameterized game indicator function library to obtain the task completion evaluation result. The pre-built parameterized game indicator function library contains multiple task evaluation indicator functions associated with specific adversarial game scenarios. The adversarial prediction module is used to input multi-source heterogeneous information into a pre-established game intent reasoning model, output intent recognition results, and associate task completion evaluation results with intent recognition results to realize the prediction and evaluation of adversarial game scenarios.

[0016] In another aspect of the present invention, in order to achieve the above-mentioned objective, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor. When the processor loads and executes the computer program, it employs the adversarial game intent prediction method based on game indicators and intent reasoning as described above.

[0017] In another aspect of the present invention, in order to achieve the above objective, a computer-readable storage medium is disclosed, wherein a computer program is stored in the computer-readable storage medium, and when the computer program is loaded and executed by a processor, it employs the adversarial game intent prediction method based on game indicators and intent reasoning as described above.

[0018] The beneficial effects of this invention are: This invention combines a parametric game index function library with a game intent reasoning model to achieve accurate modeling and evaluation of game adversarial scenarios, significantly improving the accuracy of identifying the adversary group's game intent and providing reliable multi-dimensional situation assessment and intent prediction support for agent decision-making. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the overall process of the adversarial scenario modeling and evaluation method of the present invention; Figure 3 This is a schematic diagram of the method for constructing a game indicator function library based on expert knowledge according to the present invention; Figure 4 This is a schematic diagram of the game intent reasoning process based on spatiotemporal attention and hierarchical fusion of the present invention; Figure 5 This is a schematic diagram of the system structure of the present invention; Figure 6 This is a schematic diagram of a heterogeneous multi-agent cooperative adversarial scenario according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0021] Example 1: like Figure 1 As shown, an adversarial game intent prediction method based on game indicators and intent reasoning includes the following steps: S101: Receive multi-source heterogeneous information and dynamic situational data of the game scenario, wherein the multi-source heterogeneous information includes visual modal information, non-visual temporal modal information and knowledge base modal information; Among them, visual modal information is a continuous frame image sequence that reflects the changes in the size and distribution of the adversary cluster; non-visual temporal modal information is temporal data formed by changes in the adversary's direction of travel, speed, and agent type composition; and knowledge base modal information is prior knowledge and intelligence. The continuous frame image sequence reflects the spatiotemporal evolution characteristics of the cluster, with a frame rate of M frames per second. The modal information of the knowledge base is obtained from the situational analysis knowledge base and converted into a probability vector of length k. .

[0022] S102: Map the dynamic situation data of the game scenario to a pre-built parameterized game indicator function library to obtain the task completion evaluation result. The pre-built parameterized game indicator function library contains multiple task evaluation indicator functions associated with specific adversarial game scenarios. The parameterized game index function library is constructed based on expert knowledge; The multiple task evaluation index functions associated with specific adversarial game scenarios consist of multiple parameters, including but not limited to: winning time, winning rate, and local task success rate.

[0023] Specifically, this invention constructs a parameterized game indicator function library based on expert knowledge. Addressing the problem that existing adversarial game decision-making systems use only the same indicator function to evaluate all tasks, resulting in an inability to adapt to diverse scenarios and task differences, the function library designed in this invention contains multiple game indicator functions, each associated with a specific game scenario (e.g., game scenario 1, game scenario 2... game scenario n), including but not limited to different cluster adversarial scenarios in a multi-agent game simulation platform (e.g., different terrains, different agent type configurations).

[0024] To achieve accurate evaluation of different game scenarios, this invention designs a game task evaluation index function composed of multiple parameters based on expert knowledge. These parameters reflect the quality of game task completion from multiple aspects and dimensions, and their specific parameters may include, but are not limited to, winning time, winning rate, and local task success rate. Through this parameterized design, the system can dynamically adjust the weights and parameters of each item in the game task evaluation index function according to the actual game scenario, thereby overcoming the technical bottleneck of fixed evaluation standards in traditional methods.

[0025] In its implementation, this invention designs a reasonable functional to establish a mapping relationship between the "game scenario library" and the "parametric game indicator function library." This mapping relationship is first established through simulation derivation; subsequently, expert knowledge is used to correct and optimize the functional mapping relationship. This mechanism ensures the matching degree between the game task evaluation indicator function and the specific game scenario, thereby guaranteeing the accuracy and reliability of the evaluation of the game task completion status.

[0026] S103: Input multi-source heterogeneous information into a pre-established game intent reasoning model, output the intent recognition result, and associate the task completion evaluation result with the intent recognition result to realize the prediction and evaluation of adversarial game scenarios.

[0027] The pre-established game intent reasoning model includes a visual branch module, a non-visual branch module, a knowledge base branch module, and a layered fusion module.

[0028] The visual branch module uses a 3D convolutional neural network (3D CNN) to extract spatiotemporal features from visual modal information and obtain the feature vector output by the visual branch. The non-visual branch module uses a gated recurrent unit (GRU) to extract long-term features from non-visual temporal modal information and obtain the feature vector output by the non-visual branch. The knowledge base branch module is used to process prior knowledge base modal information in the form of probability vectors to obtain the feature vector output by the knowledge base branch. The hierarchical fusion module includes a spatiotemporal attention mechanism, which performs weighted fusion on the feature vectors of the visual branch output, the non-visual branch output, and the knowledge base branch output to obtain fused features. The fused features are then input into a fully connected layer to output the game intent recognition result of the adversary cluster.

[0029] The 3D CNN convolutional kernel of the visual branch module has a × b spatial dimension and c frames of sliding time; the number of GRU hidden units in the non-visual branch module is h; the spatiotemporal attention mechanism of the hierarchical fusion module includes: a temporal attention submodule and a spatial attention submodule, which are used to assign weights to temporal evolution features and spatial distribution features, respectively.

[0030] The game intent identification result of the adversary cluster is a set of intent categories {I1, I2, …} and the probability value corresponding to each intent category, output by the fully connected layer of the hierarchical fusion module.

[0031] To address the problem of insufficient accuracy in inferring opponent intent in adversarial scenarios due to the diverse and heterogeneous sources of information on adversaries (such as quantity, state, behavior, and location) and game situation information, this invention proposes a game intent inference model based on a spatiotemporal attention mechanism and hierarchical heterogeneous information fusion. The core of this model is a multi-branch recurrent neural network structure. It uses a 3D convolutional neural network (3D CNN) and attention mechanism to spatiotemporally model the game environment and the opponent's cluster game actions, and uses a recurrent neural network (RNN) structure to hierarchically fuse heterogeneous information, ultimately achieving accurate identification of the opponent's cluster game intent.

[0032] The specific implementation of the model is as follows: First, all information available to us is decoupled and divided into three types of heterogeneous information input sources: (i) visual modal information, which refers to continuous frame image information of the opponent's cluster captured by our sensors, such as a continuous frame image sequence reflecting the spatiotemporal evolution characteristics of the opponent's cluster; (ii) non-visual temporal modal information, which refers to temporal data in the game environment situation, such as temporal data formed by changes in the opponent's direction of movement, speed, and agent type composition; and (iii) knowledge base modal information, which refers to prior knowledge and intelligence, such as knowledge from our situation analysis knowledge base, and is converted into probability vector form for input into the model. Second, the core processing module of the model is constructed, including a CNN module, an RNN module, and an attention mechanism module. Among them, the CNN module adopts a three-dimensional convolutional neural network; this module stacks the visual image information of multiple consecutive frames in the game environment in the time dimension and uses it as the input of the 3D CNN. It performs convolution operations in the time and space dimensions simultaneously through 3D convolution kernels, thereby effectively extracting the spatiotemporal features in the temporal visual images. The RNN module is composed of GRUs (Gated Recurrent Units). As a type of recurrent neural network, GRUs use internal forget, reset, and update gates to finely control the flow and forgetting of information, making them adept at extracting long-term temporal dependencies (such as non-visual information) from time-series data with significant time step distances. The attention mechanism module assigns attention weights to temporal features from different times and sources. By focusing on key features in opponent actions and changes in the game environment, the model can quickly filter high-value information, thereby improving the efficiency and accuracy of task processing.

[0033] Then, a multi-branch recurrent neural network is constructed based on the above modules to achieve hierarchical heterogeneous information fusion. This network structure includes a visual branch module, a non-visual branch module, and a knowledge base branch module: the visual branch inputs the spatiotemporal features extracted by the 3D CNN module into its dedicated GRU module and attention mechanism module to perform deep temporal fusion of the spatiotemporal features; the non-visual branch inputs the temporal data of non-visual information (such as travel speed, agent type composition, etc.) into its dedicated GRU module and attention mechanism module to extract its temporal features; and the knowledge base branch also inputs the probability vector sequence of our knowledge base into its dedicated GRU module and attention mechanism module for fusion processing.

[0034] Finally, in the branch fusion stage, the fusion feature vectors output from the three branches (visual, non-visual, and prior knowledge) are concatenated to form a high-dimensional feature vector containing all heterogeneous information. This high-dimensional vector is input into the final fully connected layer (FC) for classification prediction, thereby obtaining the recognition result of the opponent's group's game intent (e.g., intention). Figure 1 probability, meaning Figure 2 The probability of intention N is calculated as...the probability of intention N). This result provides accurate and reliable basic data support for subsequent agent decision-making. Finally, this invention correlates the intention recognition result output by the game intention reasoning model with the task completion evaluation output by the parameterized game index function library module to achieve comprehensive modeling and evaluation of this adversarial game scenario.

[0035] Specifically, the present invention will be further illustrated below through embodiments: See Figures 2-4 This invention provides a method for modeling and evaluating adversarial scenarios, the overall process of which is as follows: Figure 2 As shown: First, the system acquires two types of inputs from external sources: game scenario data and multi-source heterogeneous information (including visual, non-visual, and prior knowledge). The game scenario data is input to the parameterized game index function library module, which analyzes the scenario features and ultimately outputs the game task evaluation result. The multi-source heterogeneous information is input to the game intent inference module, which performs spatiotemporal modeling and hierarchical fusion of the information and ultimately outputs the opponent cluster's game intent. The outputs of the two modules are correlated to form the system modeling and evaluation results of this invention, providing support for subsequent multi-agent decision-making.

[0036] like Figure 3As shown, the specific implementation method of constructing the parameterized game performance indicator function library in this invention is as follows: The system first accesses a game scenario library, which contains multiple predefined adversarial scenarios (such as game scenario 1, game scenario 2... game scenario n). This scenario information is sent to a mapping module. The core function of this mapping module is to design a game task evaluation indicator function composed of multiple parameters based on expert knowledge to solve the problem that traditional evaluation indicators are singular and cannot adapt to the diversity of tasks. The parameters can reflect the quality of game task completion in multiple dimensions, for example: ax winning time + bx winning rate + cx local task success rate + ...

[0037] In this embodiment, the mapping module establishes a mapping relationship between game scenarios and game task evaluation index functions through simulation and deduction by designing reasonable functionals. Simultaneously, this mapping relationship is corrected using expert knowledge to ensure the accuracy and reliability of the evaluation. The final output of the mapping module is a parameterized game index function library, which stores parameterized index functions (e.g., index function 1, index function 2... index function n) that match each game scenario, thereby achieving dynamic evaluation and scenario adaptability.

[0038] like Figure 4 As shown, the specific implementation of the game intent reasoning process in this invention is as follows: The model receives three types of heterogeneous information inputs. In a typical heterogeneous agent cooperative game scenario, Input 1: Visual information of the adversary cluster, specifically referring to a continuous image frame sequence of the size and distribution of the adversary heterogeneous agent cluster. This type of information (such as adversary size and distribution) is essentially a continuous image frame sequence. This embodiment uses a three-dimensional convolutional neural network (3D CNN) module to process this input. 3D CNN can perform convolution operations simultaneously in the spatial dimension (width and height of the image) and the temporal dimension (continuous frames) to extract the spatiotemporal joint features of the adversary cluster over continuous time, effectively capturing its dynamic changes. Input 2: Non-visual information of the adversary cluster, specifically referring to time-series data such as the movement speed and agent type composition of the adversary heterogeneous agent cluster. This type of information is typical time-series data. This embodiment uses a gated recurrent unit (GRU) module to process this input. GRU, as an advanced recurrent neural network, possesses short-term memory capabilities and can effectively mine the temporal dynamic characteristics of such signals through its internal update and reset gates, capturing long-term temporal dependencies in opponent strategy changes. Input 3: Our prior knowledge, specifically referring to our heterogeneous agent collaborative game strategy knowledge base or situation analysis knowledge base. This type of information (such as situation analysis knowledge, expert knowledge base) is processed into a probability vector sequence. This embodiment also uses a GRU module to process this input, in order to model the evolution and correlation of prior knowledge in the time dimension.

[0039] After initial feature extraction of the three types of information mentioned above, this embodiment performs deep processing in three independent branches (visual branch, non-visual branch, and prior knowledge branch). In the visual branch, the spatiotemporal feature vector sequence extracted by the 3D CNN is fed into a dedicated GRU module to further capture higher-order temporal relationships between spatiotemporal features. Subsequently, an Attention module assigns weights to visual features at different time steps, ultimately producing the visual branch output. In the non-visual branch and the prior knowledge branch, the input data is also processed by their respective GRU and Attention modules to filter out high-value temporal information, producing the non-visual branch output and the prior knowledge branch output, respectively.

[0040] Finally, in the branch fusion and output stage, this embodiment employs a feature-level fusion strategy. The outputs of the three branches (visual branch output, non-visual branch output, and prior knowledge branch output) are concatenated and fused to form a high-dimensional unified feature vector containing all heterogeneous information. This vector then passes through a top-level Attention module, which adaptively calculates the contribution weights of the three branches, enabling the model to dynamically focus on the most important information source based on the current situation (e.g., focusing more on visual information when rapidly approaching an opponent, and more on prior knowledge during the latent phase). The weighted fused feature vector is then fed into a fully connected layer as a classifier, ultimately outputting the adversary's cluster intent probability vector, i.e., the probability of the adversary's intention (e.g., intention). Figure 1 ,meaning Figure 2 The recognition result of ...). The final step of this invention, as... Figure 1 As shown, the game intent reasoning module ( Figure 4 The intent recognition result output by the parameterized game index function library module () Figure 3 The task completion evaluation output is correlated to achieve system modeling and evaluation of this adversarial game scenario.

[0041] Example 2: To further verify the effectiveness of the adversarial game intent prediction method based on game indicators and intent reasoning proposed in this invention in complex dynamic environments, this example constructs a heterogeneous multi-agent cooperative adversarial simulation scenario, and demonstrates the operation process and exemplary output data of the method in a specific scenario based on the method flow described in Example 1.

[0042] The system first acquires two types of input from the external simulation environment: game scenario data and multi-source heterogeneous information. In this embodiment, a heterogeneous multi-agent cooperative adversarial scenario (such as...) is constructed. Figure 6 As shown in the image, the scene includes an adversary cluster on the left and a friendly cluster on the right. The adversary consists of three heterogeneous intelligent agents: Type A basic firing units located on the front ground (corresponding to...). Figure 6The smaller ground unit on the left side of the middle row), the Type B heavy assault unit located in the middle row of the ground (corresponding to...) Figure 6 The larger ground unit on the left and the type C aerial execution unit located in the air behind (corresponding to) Figure 6 (The flying unit hovering on the left side of the middle). Figure 6 It visually demonstrates that the opposing group is advancing towards our area on the right and launching an attack.

[0043] The specific multi-source heterogeneous information received by the system includes: (i) visual modal information: receiving feature map sequences output by the simulation environment, with the frame rate set to M=30 frames / second. This image sequence shows that the adversary cluster is densely distributed, with ground execution units in front and air execution units hovering behind to provide support, presenting an overall "arc-shaped encirclement" trend advancing towards our area on the right (spatiotemporal evolution characteristics). Figure 6 As shown, the visual information clearly captures the key action features of the Type C airborne execution unit releasing energy beams to repair the front-row ground execution units. (ii) Non-visual temporal modal information: Receive state temporal data uploaded by each agent. Simulation results show that the average movement speed of the adversary Type A and Type B ground execution units drops to 0 m / s (indicating that they have entered a contact confrontation state), and the Type C airborne execution unit is releasing "energy transfer / repair" signals at high frequency. The composition ratio of agent types is "7:2:1" (basic: heavy-armored: airborne). (iii) Knowledge base modal information: Load the "heterogeneous multi-agent cooperative strategy knowledge base", extract the strategy rules related to the current close contact situation, and convert them into probability vectors of length k=64 for input into the model.

[0044] The construction and application of the parameterized game indicator function library in this invention are as follows: The system inputs the dynamic situation data of the above-mentioned game scenario into the parameterized game indicator function library module. The module identifies the current scenario characteristics (heterogeneous formation, air support, high-intensity contact) and automatically matches the "heterogeneous cluster confrontation evaluation indicator function" through a preset functional. This module uses the pre-constructed functional mapping relationship to comprehensively calculate the input dynamic situation data (such as the distance between the two sides and terrain factors), and combines it with historical statistical data from simulation to output the quantitative evaluation indicators of the adversary's task in the current scenario as follows: the estimated task completion time is 45 seconds (evaluated as "fast"); the estimated advantage probability (win rate) is 62.0% (evaluated as "slightly better than the mean"); the local task (unit survival rate) success rate is 85.0% (thanks to the continuous repair effect of the type C air execution unit).

[0045] The specific process of game intent reasoning in this invention is as follows: After receiving visual, non-visual, and knowledge base modal information, the computer system extracts the spatial distribution features of the adversary cluster's "front row withstands pressure, back row outputs" in the visual branch using 3D CNN; in the non-visual branch, GRU extracts the long-term temporal dependencies of the adversary's skill release frequency surge. In the hierarchical fusion module, the position information of "Type C aerial execution unit" is given the highest weight (weight value 0.65) through a spatiotemporal attention mechanism, because it is a key node for maintaining the formation's sustainability. In the branch fusion and output stage, the model outputs the following example of the adversary cluster's game intent probability vector: {Intent I1: Concentrated collaborative confrontation, probability: 0.89}, {Intent I2: Pulling and evading, probability: 0.07}, {Intent I3: Dispersed withdrawal, probability: 0.04}. The system determines that the adversary's current main intent is "concentrated collaborative confrontation".

[0046] In the final step of this invention, the intent recognition result ("centralized collaborative confrontation", probability 0.89) output by the game intent reasoning module is correlated with the task completion evaluation ("advantage probability 62.0%) output by the parameterized game index function library module to generate the final system modeling and evaluation report: "The current adversary is implementing a centralized collaborative confrontation strategy (high confidence), and due to the effective operation of its air execution unit, its local task success rate is relatively high. It is recommended that we prioritize interfering with or suppressing its type C air execution unit to disrupt its collaborative system." Example 3: To achieve the above objective, such as Figure 5 As shown, based on Embodiment 1, this invention discloses an adversarial game intent prediction system based on game indicators and intent reasoning, comprising: The data receiving module 11 is used to receive multi-source heterogeneous information and dynamic situational data of the game scenario. The multi-source heterogeneous information includes visual modal information, non-visual temporal modal information and knowledge base modal information. Among them, visual modal information is a continuous frame image sequence that reflects the changes in the size and distribution of the adversary cluster; non-visual temporal modal information is temporal data formed by changes in the adversary's direction of travel, speed, and agent type composition; and knowledge base modal information is prior knowledge and intelligence. The relation mapping module 12 is used to map the dynamic situation data of the game scenario to a pre-built parameterized game index function library to obtain the task completion evaluation result. The pre-built parameterized game index function library contains multiple task evaluation index functions associated with specific adversarial game scenarios. The adversarial prediction module 13 is used to input multi-source heterogeneous information into a pre-established game intent reasoning model, output the intent recognition result, and associate the task completion evaluation result with the intent recognition result to realize the prediction and evaluation of adversarial game scenarios.

[0047] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., serving as the computing and control core of the terminal, and is used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the aforementioned method.

[0048] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0049] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0050] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A method for predicting intentions in adversarial games based on game indicators and intention reasoning, characterized in that, The method includes the following steps: Receive multi-source heterogeneous information and dynamic situational data of the game scenario, wherein the multi-source heterogeneous information includes visual modal information, non-visual temporal modal information and knowledge base modal information; Among them, visual modal information is a continuous frame image sequence that reflects the changes in the size and distribution of the adversary cluster; non-visual temporal modal information is temporal data formed by changes in the adversary's direction of travel, speed, and agent type composition; and knowledge base modal information is prior knowledge and intelligence. The dynamic situational data of the game scenario is mapped to a pre-built parameterized game indicator function library to obtain the task completion evaluation result. The pre-built parameterized game indicator function library contains multiple task evaluation indicator functions associated with specific adversarial game scenarios. By inputting multi-source heterogeneous information into a pre-established game intent reasoning model, the intention recognition result is output. The task completion evaluation result is correlated with the intention recognition result, thus realizing the prediction and evaluation of adversarial game scenarios.

2. The adversarial game intent prediction method based on game indicators and intent reasoning according to claim 1, characterized in that, The continuous frame image sequence reflects the spatiotemporal evolution characteristics of the cluster, with a frame rate of M frames per second. The modal information of the knowledge base is obtained from the situational analysis knowledge base and converted into a probability vector of length k. .

3. The adversarial game intent prediction method based on game indicators and intent reasoning according to claim 1, characterized in that, The parameterized game index function library is constructed based on expert knowledge; The multiple task evaluation index functions associated with specific adversarial game scenarios consist of multiple parameters, including but not limited to: winning time, winning rate, and local task success rate.

4. The adversarial game intent prediction method based on game indicators and intent reasoning according to claim 1, characterized in that, The pre-established game intent reasoning model includes a visual branch module, a non-visual branch module, a knowledge base branch module, and a layered fusion module.

5. The adversarial game intent prediction method based on game indicators and intent reasoning according to claim 4, characterized in that, The visual branch module uses a three-dimensional convolutional neural network (3D CNN) to extract spatiotemporal features from visual modal information and obtain the feature vector output by the visual branch. The non-visual branch module uses a gated recurrent unit (GRU) to extract long-term features from non-visual temporal modal information and obtain the feature vector output by the non-visual branch. The knowledge base branch module is used to process prior knowledge base modal information in the form of probability vectors to obtain the feature vector output by the knowledge base branch. The hierarchical fusion module includes a spatiotemporal attention mechanism, which performs weighted fusion on the feature vectors of the visual branch output, the non-visual branch output, and the knowledge base branch output to obtain fused features. The fused features are then input into a fully connected layer to output the game intent recognition result of the adversary cluster.

6. The adversarial game intent prediction method based on game indicators and intent reasoning according to claim 5, characterized in that, The 3D CNN convolution kernel of the visual branch module slides in a spatial dimension of a × b and a temporal dimension of c frames. The number of GRU hidden units in the non-visual branch module is h; The spatiotemporal attention mechanism of the hierarchical fusion module includes a temporal attention submodule and a spatial attention submodule, which are used to assign weights to temporal evolution features and spatial distribution features, respectively.

7. The adversarial game intent prediction method based on game indicators and intent reasoning according to claim 6, characterized in that, The game intent identification result of the adversary cluster is a set of intent categories {I1, I2, …} and the probability value corresponding to each intent category, output by the fully connected layer of the hierarchical fusion module.

8. A system for predicting adversarial game intent based on game indicators and intent reasoning, employing the adversarial game intent prediction method based on game indicators and intent reasoning as described in any one of claims 1 to 7, characterized in that, include: The data receiving module is used to receive multi-source heterogeneous information and dynamic situational data of the game scenario. The multi-source heterogeneous information includes visual modal information, non-visual temporal modal information and knowledge base modal information. Among them, visual modal information is a continuous frame image sequence that reflects the changes in the size and distribution of the adversary cluster; non-visual temporal modal information is temporal data formed by changes in the adversary's direction of travel, speed, and agent type composition; and knowledge base modal information is prior knowledge and intelligence. The relation mapping module is used to map the dynamic situation data of the game scenario to a pre-built parameterized game indicator function library to obtain the task completion evaluation result. The pre-built parameterized game indicator function library contains multiple task evaluation indicator functions associated with specific adversarial game scenarios. The adversarial prediction module is used to input multi-source heterogeneous information into a pre-established game intent reasoning model, output intent recognition results, and associate task completion evaluation results with intent recognition results to realize the prediction and evaluation of adversarial game scenarios.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs the adversarial game intent prediction method based on game indicators and intent reasoning, as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it employs the adversarial game intent prediction method based on game indicators and intent reasoning as described in any one of claims 1 to 7.

Images