A battlefield key situation extraction method and system based on a war game system

Abstract

The application discloses a battlefield key situation extraction method and system based on a war game system, which comprises the following steps: obtaining original situation data and inputting the original situation data into a situation screening model to obtain screened situation data; inputting the original situation data and the screened situation data into an intelligent agent decision model respectively to obtain an original intelligent agent action space and a screened intelligent agent action space; calculating a loss function based on the original intelligent agent action space and the screened intelligent agent action space, and updating the situation screening model according to the calculated loss value, and repeating the training until a training end condition is reached; processing the original situation data to be screened by using the trained situation screening model to output the screened situation data, and completing the battlefield key situation extraction. The application is used for assisting in explaining simulation deduction results, solving the problem of key situation data identification and extraction, opening the black box of the intelligent agent decision model in the deduction system, and being beneficial to the correlation degree analysis between the deduction results and the situation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of situation screening, and particularly relates to a battlefield key situation extraction method and system based on a war game system. BACKGROUND

[0002] The original battlefield situation information in the war game system is mainly characterized by high information dimension, multiple features and nonlinearity, and the situation screening only relying on the expert experience of the operator has strong subjectivity, and reasonable and key situation information is a key condition for the intelligent agent to make correct decisions, so a method for extracting key situation information from complex situation information is needed to assist in explaining the deduction result. In the complex battlefield environment, the battlefield key situation extraction method is used for identifying and extracting key situation and the interpretability of key situation information, opening the black box of the intelligent agent decision model in the deduction system, and facilitating the correlation analysis between the deduction result and the situation.

[0003] At present, the common key situation data extraction method based on mathematical model has certain limitations in processing high-dimensional features and nonlinear data, resulting in inaccurate key information screened out; and in the situation screening process using the method of machine learning, as the feature dimension is reduced, the lost feature information also increases, which also affects the effectiveness of key situation extraction.

[0004] Correspondingly, for the battlefield key situation extraction method based on the war game system, the existing technologies such as "Multi-level Situation Generation Method Based on Massive Joint Battlefield Information" (author Wang Hongjie, Feng Yanlai), "Algorithm Research on Battlefield Situation Target Recognition and Situation Intention Prediction" (author Wang Yang), "Battlefield Situation Analysis and Prediction Based on Big Data Analysis Technology" (author Wang Xiujun, Cao Jin, Wang Jianqiang, Han Wenhua), "Sea Battlefield Situation Analysis and Evaluation Method" (author Wang Xiaofei, Chen Yunqiu) and the like. However, in the situation extraction based on the war game system, the following deficiencies exist at present: the situation screening technology based on deep learning algorithm has large feature loss, the screening process is a black box, and the interpretability is poor; the situation screening technology based on mathematical model has great limitations in processing high-dimensional features and nonlinear original situation data. SUMMARY

[0005] One of the purposes of the present application is to provide a battlefield key situation extraction method based on a war game system, which is used for assisting in explaining the simulation deduction result, solving the problem of key situation data identification and extraction, opening the black box of the intelligent agent decision model in the deduction system, and facilitating the correlation analysis between the deduction result and the situation.

[0006] To achieve the above purpose, the technical solution adopted by the present application is as follows:

[0007] A method for extracting key battlefield situations based on a war game simulation system, comprising:

[0008] Obtaining original situation data and inputting it into the situation screening model to obtain the filtered situation data;

[0009] The original situation data and the filtered situation data are input into the agent decision model respectively to obtain the original agent action space and the filtered agent action space;

[0010] Calculating a loss function based on the original agent action space and the filtered agent action space, and updating the situation screening model according to the calculated loss value, and repeating the training until the training end condition is met;

[0011] The trained situation screening model is used to process the original situation data to be screened, and the screened situation data is output to complete the extraction of key battlefield situations.

[0012] Several optional methods are also provided below, but they are not intended to be additional limitations on the above-mentioned overall solution. They are merely further supplements or optimizations. Under the premise that there are no technical or logical contradictions, each optional method can be combined separately for the above-mentioned overall solution, or multiple optional methods can be combined.

[0013] Preferably, the agent decision model is constructed based on the DQN algorithm.

[0014] Preferably, when the war game simulation system performs a drone strike mission, the process of constructing the agent decision model is as follows:

[0015] Definition of state space: select aircraft speed, aircraft position, missile state, number of missiles, missile flight time, aircraft mount state, target position, and target current speed;

[0016] Definition of action space: assuming our drone is a mobile target, the set of actions available to our drone includes firing and moving; assuming the enemy target is a fixed target, the set of actions available to the enemy target includes firing;

[0017] Definition of reward function: The reward of the agent is represented by R, which is as follows:

[0018]

[0019] C=∑ i C i

[0020] W=∑ j W j

[0021]

[0022]

[0023] wherein, W is the damage value of the enemy target, C is the cost of the UAV of our side, i represents the batch number of the UAV of our side, j represents the batch number of the enemy target, i, j = 1, 2, 3, …, C i represents the cost of the UAV of our side with the batch number i, W j represents the damage value of the target of the enemy side with the batch number j.

[0024] As preferred, the original situation data is time series information {X t |t∈{1,2,…,T}}, wherein, is the battlefield situation data collected from the war game system at time t and selected through expert experience, is the n th battlefield situation data in X t , n∈{1,2,…,N}, T is the time step.

[0025] As preferred, the situation screening model comprises an encoding layer, an attention layer and a situation fusion layer;

[0026] The encoding layer comprises T gated recurrent units, and each gated recurrent unit maps the battlefield situation data X t input at the corresponding time into a hidden state value h t .

[0027] The attention layer aggregates the hidden state value h t output by the encoding layer and the corresponding battlefield situation data X t , calculates the similarity between the aggregated situation data and the battlefield situation data, obtains a similarity matrix composed of nT similarity values, and performs normalization processing on the similarity matrix to obtain an importance mask corresponding to each battlefield situation data, thereby generating an attention importance mask matrix;

[0028] The situation fusion layer performs matrix point multiplication on the importance mask matrix and the original situation data, and outputs the screened situation data.

[0029] As preferred, the attention layer aggregates the hidden state value h t output by the encoding layer and the corresponding battlefield situation data X t , and calculates the similarity between the aggregated situation data and the battlefield situation data, including:

[0030]

[0031]

[0032]

[0033] Where S t is the hidden state value h t Corresponding battlefield situation data X t Aggregated situation data after aggregation, x t ′n is the nth aggregated battlefield situation data in the aggregated situation data, ⊙ represents the Hadamard product, s tn Represents the nth battlefield situation data at time t and the nth aggregated battlefield situation data x at time t t ′n The similarity, Represents the similarity matrix.

[0034] Preferably, the similarity matrix is ​​normalized to obtain an importance mask corresponding to each battlefield situation data, and an attention importance mask matrix is ​​generated, including:

[0035]

[0036]

[0037] Where, Similarity s tn The normalized value is the importance mask of the nth battlefield situation data at time t. is the attention importance mask matrix.

[0038] Preferably, the loss function is calculated based on the original agent action space and the filtered agent action space, including: minimizing the loss L mask As a loss function, the expression is as follows:

[0039]

[0040] Where M is the batch size of training samples, D i ′ is the action space of the agent after screening corresponding to the i-th original situation data, D i is the original agent action space corresponding to the i-th original situation data, α is the adjustment coefficient, is the attention importance mask matrix corresponding to the i-th original situation data, and ‖·‖1 represents the L1 norm.

[0041] This invention provides a method for extracting critical battlefield situations based on a wargaming system. Through self-supervision, an intelligent agent can independently select valid, mission-related situational information. This addresses the current problem of relying primarily on operators' own empirical estimates to identify critical situations, resulting in low efficiency and difficulty in accurately extracting important situational information. This embodiment quickly and accurately screens critical situations, unlocking the black box of the agent's decision-making model, facilitating the interpretation and analysis of game results and improving the interpretability of the agent.

[0042] A second object of the present invention is to provide a battlefield key situation extraction system based on a war game simulation system, comprising a processor and a memory storing a plurality of computer instructions, wherein the computer instructions, when executed by the processor, implement the steps of the battlefield key situation extraction method based on the war game simulation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is an architecture diagram of a battlefield key situation extraction system method based on a war game simulation system of the present invention;

[0044] Figure 2 Schematic diagram of the structure of the situation screening model of the present invention;

[0045] Figure 3 Schematic diagram of the structure of the gated recurrent unit of the present invention;

[0046] Figure 4 This is a training flow chart of the situation screening model of the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0049] In order to overcome the defects of situation extraction in the prior art, this embodiment proposes a battlefield key situation extraction method based on a war game simulation system, such as Figure 1 As shown in the figure, this method is mainly implemented based on the situation screening model and the intelligent agent decision model.

[0050] The agent decision model is constructed based on a reinforcement learning algorithm, such as the DQN algorithm, the PPO algorithm, the DDPG algorithm, etc. The DQN algorithm is preferred in this embodiment, and other algorithms can be selected as needed in other embodiments.

[0051] The process of building an agent decision model varies depending on the reinforcement learning algorithm. This example uses the DQN algorithm as an example to illustrate the construction process. The construction process primarily involves defining the state space, action space, and reward function. For ease of understanding, the application scenario of the agent decision model is set to a naval battle scenario involving drone strike missions. As the application scenario changes, the corresponding defined state space, action space, and reward function can also change accordingly.

[0052] Definition of state space: Select aircraft speed, aircraft position (latitude and longitude, altitude), missile status, number of missiles, missile release angle, missile flight time, aircraft detection radar status, aircraft mounting status, target position, target current speed, target acceleration, etc.

[0053] Action Space Definition: Since our drones are mobile targets, the action set for all our aircraft includes firing and moving. Our drones fire based on radar detection, and they move at every moment. Enemy maritime targets are fixed targets, and the enemy's action set is firing. They fire based on radar detection and immediately fire upon detecting our aircraft.

[0054] Definition of reward function: The reward of the agent is represented by R, which is as follows:

[0055] R=WC

[0056] C=∑ i C i

[0057] W=∑ j W j

[0058]

[0059]

[0060] Where W is the damage value of the enemy target, C is the cost of our UAV, i represents the batch number of our UAV, j represents the batch number of the enemy target, i, j = 1, 2, 3, ..., C i represents the cost of our drone with batch number i, W j Indicates the damage value of the enemy target with batch number j.

[0061] In addition, if Figure 2As shown, the situation screening model established in this embodiment includes an input layer, an encoding layer, an attention layer, and a situation fusion layer. The key of the model is to use n situation data in T time slices to screen out effective situation data based on a gated recurrent unit (GRU) and an attention mechanism. The attention mechanism generates an importance mask matrix for the data group decoded by the GRU, which is used to filter relatively unimportant situation data.

[0062] (1) Input layer: The input of the input layer is original situation data.

[0063] The original situation data of this embodiment is time series information {X t |t∈{1,2,…,T}}. Among them, is the battlefield situation data selected by experts from the war game system at t time, is the n th battlefield situation data in X t , n∈{1,2,…,N}, and T is the time step. That is, the input is

[0064] This embodiment takes the data in one time step as input, and the battlefield situation data at times 1 to T-1 is used to assist in obtaining the correlation between each battlefield situation data at time T.

[0065] And in order to make the network model more easily identify the input quantity representing the situation, the original situation data transmitted to the input layer is preprocessed data, and the preprocessing here can be normalization, onehot encoding, binary encoding, etc. For example, for floating-point numbers such as aircraft speed, flight time, etc., directly normalize the unit information type, and unify the information of different units and different measurement standards to a quantity between 0 and 1; for enumerable type numbers such as missile state, aircraft load state, onehot encode the integer representing the type; for dynamically changing information such as aircraft coordinate changes, perform binary encoding.

[0066] (2) Encoding layer: Select the gated recurrent unit as the encoder to perform data dimensionality reduction and feature extraction on the original battlefield data of different time periods obtained from the war game system, and capture the dependency relationship between the situation data of each time period.

[0067] The encoding layer is composed of a gated recurrent unit, and the internal structure of the GRU is as shown in Figure 3 . By learning the flow of "gate" control information, the dependency relationship between data with large time step distances in the input time series is better captured. The GRU as the neural network processing unit of the encoder obtains data from the input layer, and the output hidden state value provides data support for the encoding of the attention layer.

[0068] Initialize the GRU neural network parameters and map the input time series information to the hidden state of the encoder according to the pre-established GRU encoding layer. In order to save resources, this embodiment sets the encoding layer to include T gated recurrent units. Each gated recurrent unit maps the battlefield situation data X input at the corresponding moment to the hidden state of the encoder. t Mapped to hidden state value h t Specific mapping process In this embodiment, the t-th gated recurrent unit processes the battlefield situation data X at time t. t Take this as an example to illustrate.

[0069] Battlefield situation dataX t-1 After being processed by the previous gated recurrent unit, its output hidden state h t-1 Will be passed to the current gated recurrent unit. The updated gate state z of the current gated recurrent unit t and reset gate state r t They are:

[0070] r t =σ(w r .[h t-1 ,X t ])

[0071] z t =σ(w z .[h t-1 ,X t ])

[0072] Where σ is the sigmoid function, which transforms the data into a value in the range of 0-1, thus acting as a gating signal, and w r To reset the gate weight, w z is the weight of the update gate.

[0073] Data reset process: After obtaining the gate signal, first use the reset gate to obtain the data h′ after “reset” t-1 =r t *h t-1 , and then h′ t-1 With X t Splicing is performed, and the data is scaled to the range of -1 to 1 through the tanh activation function, which is obtained as follows:

[0074]

[0075] Where, Reset the weights of the process for the data.

[0076] Update process: Two steps of selective forgetting and memory are performed simultaneously:

[0077]

[0078] In the formula, (1-z t )*h t-1 Indicates the selective forgetting of the original hidden state information, (1-z t ) can be seen as a forget gate, Indicates the retention of some past memories, so the information output by the current gated recurrent unit is h t In the gated recurrent unit, z t To control the update gate, it is used to update the current encoder to retain the previous state information h t-1 The larger the value of the update gate is, the more information is brought in at the previous moment; t To control the reset gate, determine the current encoder to forget the historical information h t-1 The smaller the reset gate is, the less information of the previous state is written.

[0079] To filter out important situations, the attention layer effectively focuses on the encoder output h t ,Right now:

[0080]

[0081] In the formula, GRU is a gated recurrent unit.

[0082] (3) Attention layer: Combined with the attention mechanism, each state data corresponds to an explanatory importance mask. The attention mechanism can process the information obtained by the encoding layer from the input original situation data, mainly solving the representation problem. By aggregating data with large differences, comparing the similarity between the aggregated situation data and the original situation data, quantitatively assigning different weights to each state data, and then generating the attention importance mask matrix of the situation data This is beneficial for the subsequent screening of situation data. This process takes the output of the gated recurrent unit and the original situation data as input, with the goal of performing differential data fusion on the original situation data. It is specifically divided into two steps.

[0083] (3.1) Calculate the similarity between the situation data output by the gated recurrent unit and the original input data.

[0084] The attention layer converts the hidden state value h output by the encoding layer into y Corresponding battlefield situation data X t Aggregation is performed, similarity is calculated, and the similarity between the aggregated situation data and the battlefield situation data is calculated to obtain a similarity matrix composed of nT similarity values, and then the correlation between the data is detected.

[0085]

[0086]

[0087]

[0088] In the formula, S y is a hidden state value h t and the corresponding battlefield situation data X y aggregated situation data x′ t n is the nth aggregated battlefield situation data in the aggregated situation data, ⊙ represents the Hadamard product, s tn represents the nth battlefield situation data at time t and the nth aggregated battlefield situation data x′ t n the similarity of represents the similarity matrix.

[0089] (3.2) The similarity matrix is normalized to obtain an importance mask corresponding to each battlefield situation data, and the sum of the importance masks corresponding to the n battlefield situation data at the same time is ensured to be 1, and an attention importance mask matrix is generated. According to the importance mask, the weight of each state data input is known, that is, the weight of the effective data is larger.

[0090]

[0091]

[0092] In the formula, is the similarity s tn the normalized value, that is, the importance mask of the nth battlefield situation data at time t is the attention importance mask matrix.

[0093] (4) Situation fusion layer: The situation fusion layer performs matrix point multiplication on the importance mask matrix and the original situation data to fuse the situation data and output the screened situation data. Through situation fusion, effective situations are screened out, while relatively unimportant situation data is filtered out, and the output dimension of this layer is consistent with the dimension of the original situation data.

[0094] In this embodiment, the battlefield situation data of one time step T is taken as a training sample, and the training sample is input into the situation screening model to obtain the screened situation data and an attention importance mask matrix corresponding thereto. In the training process, there are multiple training samples, so steps (1) to (4) are repeated to obtain the screened situation data and the attention importance mask matrix corresponding to each training sample.

[0095] The filtered situation data and attention importance mask matrix obtained in this embodiment are mainly used to train and update the situation screening model. Therefore, before updating the situation screening model, the agent decision model is first trained using training samples. When the agent decision model is trained to convergence, the combined situation screening model and the agent decision model are used to complete the training and update of the situation screening model. The parameters of the agent decision model remain fixed during the training and update of the situation screening model.

[0096] like Figure 4 As shown, based on the constructed situation screening model and agent decision model, this embodiment provides a battlefield key situation extraction method based on a war game simulation system. The specific steps are as follows:

[0097] Step 1: Obtain the original situation data and input it into the situation screening model to obtain the filtered situation data.

[0098] Step 2: Input the original situation data and the filtered situation data into the agent decision model respectively to obtain the original agent action space and the filtered agent action space.

[0099] It should be noted that the original situation data can be input into the intelligent agent decision model at the same time as the filtered situation data in step 2, or the original situation data can be obtained in step 1 and then input into the intelligent agent decision model to obtain the original intelligent agent action space as a label. There is no restriction on the order of input.

[0100] Step 3: Calculate the loss function based on the original agent action space and the filtered agent action space, and update the situation screening model according to the calculated loss value. Repeat the training until the training end condition is reached.

[0101] The loss function proposed in this embodiment contains two constraints: one is to maximize the similarity of agent behavior, and the other is to filter out as much useless situation data as possible. The optimization goal is to minimize the difference between the action space of the filtered agent and the original agent, specifically to minimize the loss L mask As a loss function, the expression is as follows:

[0102]

[0103] Where M is the batch size of the training sample, D i ′ is the action space of the agent after screening corresponding to the i-th original situation data, D i is the original agent action space corresponding to the i-th original situation data, α is the adjustment coefficient, which is a positive scalar that controls the sparsity of the importance mask. is the attention importance mask matrix corresponding to the i-th original situation data, and ‖·‖1 represents the L1 norm. Used to compare the changes in agent behavior before and after situation screening; It is used to achieve the sparseness of situation data features, and adjust the situation data that the agent pays attention to by adjusting the coefficient α, thereby filtering out invalid situation information.

[0104] It should be noted that the adjustment coefficient α can be input based on experience before each model training, or can be set to automatically adjust according to predefined rules, which is not limited in this embodiment.

[0105] Step 4: Use the trained situation screening model to process the raw situation data to be screened, outputting the filtered situation data to complete the extraction of key battlefield situations. This step belongs to the inference and application phase of the situation screening model. It is included in Step 4 to indicate that the situation screening model must be trained before application, not to restrict the application of the situation screening model to the first three steps.

[0106] The present invention provides a battlefield key situation screening method based on war game simulation. Compared with the existing deep learning algorithm-based scheme that requires clear data labels, the present method combines deep learning and reinforcement learning algorithms to first obtain the action space and then train a self-supervised situation screening model. The advantages of this scheme are: (1) combining deep learning and reinforcement learning methods, constructing a self-supervised situation screening model for screening high-dimensional, multi-feature, and nonlinear situation data, without relying on expert experience for subjective situation screening; (2) the intelligent agent decision model and the situation screening model are independent of each other, but are connected when making decisions, forming a self-supervised situation screening model; (3) combining the attention mechanism, generating a situation attention importance mask matrix, the mask matrix can be adjusted by adjusting the coefficient, and performing dot multiplication with the original situation data to perform situation data fusion.

[0107] In another embodiment, the present invention also provides a battlefield key situation extraction system based on a war game simulation system, including a processor and a memory storing a plurality of computer instructions, wherein the computer instructions, when executed by the processor, implement the steps of a battlefield key situation extraction method based on a war game simulation system.

[0108] Regarding the specific limitations of the battlefield key situation extraction system based on the war game simulation system, please refer to the limitations of the battlefield key situation extraction method based on the war game simulation system above, which will not be repeated here.

[0109] The memory and the processor are electrically connected directly or indirectly to realize the transmission or interaction of data. For example, the elements can be electrically connected through one or more communication buses or signal lines. The memory stores a computer program executable on the processor. The processor realizes the method in the embodiments of the application by running the computer program stored in the memory.

[0110] The memory can be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. The memory is used to store a program. The processor executes the program after receiving an execution instruction.

[0111] The processor can be an integrated circuit chip with data processing capability. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc. The processor can realize or execute the methods, steps and logic block diagrams disclosed in the embodiments of the application. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0112] The technical features of the above-described embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the technical features in the above-described embodiments are not described, however, as long as the combinations of the technical features do not exist in contradiction, it should be considered that the combinations are within the scope of the present disclosure.

[0113] The above-described embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, however, it should not be understood as a limitation on the scope of the application. It should be pointed out that, for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for extracting key battlefield situations based on a war game simulation system, characterized in that: The method for extracting key battlefield situations based on a war game simulation system includes: Obtain the original situation data and input it into the situation screening model to obtain the filtered situation data; the original situation data is time series information ,in, for The battlefield situation data collected from the war game simulation system at all times and selected by experts, for The Battlefield situation data, , is the time step; The original situation data and the filtered situation data are input into the agent decision model respectively to obtain the original agent action space and the filtered agent action space; Calculating a loss function based on the original agent action space and the filtered agent action space, and updating the situation screening model according to the calculated loss value, and repeating the training until the training end condition is met; The trained situation screening model is used to process the original situation data to be screened, and the screened situation data is output to complete the extraction of key battlefield situations. The situation screening model includes an encoding layer, an attention layer and a situation fusion layer; The coding layer includes Gated recurrent units, each of which converts the battlefield situation data input at the corresponding moment into Mapped to hidden state value ; The attention layer converts the hidden state value output by the encoding layer into Corresponding battlefield situation data Aggregate and calculate the similarity between the aggregated situation data and the battlefield situation data to obtain Similarity values ​​are normalized to obtain the importance mask corresponding to each battlefield situation data, and the attention importance mask matrix is ​​generated; The situation fusion layer performs matrix dot multiplication on the importance mask matrix and the original situation data, and outputs the filtered situation data.

2. The method for extracting key battlefield situations based on a war game simulation system according to claim 1, wherein: The agent decision model is built based on the DQN algorithm.

3. The method for extracting key battlefield situations based on a war game simulation system according to claim 2, wherein: When the war game simulation system performs a drone strike mission, the process of constructing the agent decision model is as follows: Definition of state space: select aircraft speed, aircraft position, missile state, number of missiles, missile flight time, aircraft mount state, target position, and target current speed; Definition of action space: assuming our drone is a mobile target, the set of actions available to our drone includes firing and moving; assuming the enemy target is a fixed target, the set of actions available to the enemy target includes firing; Definition of reward function: The reward function of the agent is Indicates that: ； ； ； ； ； Where, is the damage value of the enemy target, The cost of our drone, Indicates the batch number of our drone, Indicates the batch number of the enemy target, , Indicates our batch number is The cost of drones, Indicates that the enemy batch number is The damage value of the target.

4. The method for extracting key battlefield situations based on a war game simulation system according to claim 1, wherein: The attention layer converts the hidden state value output by the encoding layer into Corresponding battlefield situation data Perform aggregation and calculate the similarity between the aggregated situation data and the battlefield situation data, including: ； ； ； Where, is the hidden state value Corresponding battlefield situation data Aggregated situation data after aggregation, The first Aggregate battlefield situation data, represents the Hadamard product, express The moment Battlefield situation data and The moment Aggregated battlefield situation data The similarity, Represents the similarity matrix.

5. The method for extracting key battlefield situations based on a war game simulation system according to claim 4, wherein: The similarity matrix is ​​normalized to obtain the importance mask corresponding to each battlefield situation data, and the attention importance mask matrix is ​​generated, including: ； ； Where, For similarity The normalized value is The moment The importance mask of battlefield situation data, is the attention importance mask matrix.

6. The method for extracting key battlefield situations based on a war game simulation system according to claim 1, wherein: The loss function is calculated based on the original agent action space and the filtered agent action space, including: minimizing the loss As a loss function, the expression is as follows: ； Where, is the batch size of training samples, For the The filtered agent action space corresponding to the original situation data, For the The original agent action space corresponding to the original situation data, is the adjustment coefficient, For the The attention importance mask matrix corresponding to the original situation data, express norm.

7. A battlefield key situation extraction system based on a war game simulation system, comprising a processor and a memory storing a plurality of computer instructions, characterized in that: When the computer instructions are executed by the processor, the steps of the battlefield key situation extraction method based on the war game simulation system as described in any one of claims 1 to 6 are implemented.

Images