A micro-differential game strategy-oriented intelligent control law inversion method
Patent Information
- Application Number
- CN202410022258.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-05
- Publication Date
- 2026-10-09
- Estimated Expiration
- 2044-01-05
AI Technical Summary
[0003]本发明的目的在于解决现有技术中对目标控制策略进行识别和反演时采用传统算法耗时长的技术问题，提供一种面向微分博弈策略的智能控制律反演方法
[0042] This invention discloses an intelligent control law inversion method for differential game strategies, proposing a linear time-invariant control matrix inversion method based on an intelligent learning structure using a neural network model. By constructing a neural network deep learning structure suitable for sequence data processing, it aims to more flexibly handle the complexity of linear time-invariant feedback control. Taking the linear feedback control law under differential games as the object, the method reveals the opponent's control strategy by inverting the unknown weight matrix and control law parameters, providing theoretical support for effectively responding to continuous thrust pursuit.
Smart Images
Figure CN117850232B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of spacecraft game theory technology, and relates to an intelligent control law inversion method for differential game strategies. Background Technology
[0002] In real-world game scenarios, due to the inability to directly obtain the opponent's true purpose, their maneuverability, and accurate relative state information, decision-making regarding non-cooperative targets often faces the dilemma of incomplete information. Researching efficient and intelligent strategy inversion methods to address issues such as strategy identification in non-cooperative target games is crucial. Among these methods, maneuver detection for non-cooperative targets is a prerequisite and important guarantee for control strategy inversion, providing further data support for subsequent strategy inversion. In their paper "An escape strategy in orbital pursuit-evasion games with incomplete information," Li et al. designed an effective escape strategy to address orbital pursuit-evasion games under incomplete information using the unscented Kalman online information estimation method and optimal response optimal control. In their paper "Incomplete Information Pursuit-Evasion Game Control for a Space Non-Cooperative Target," Wang et al. proposed a method to treat unknown target maneuvers as colored noise, thus reducing the game to a strong tracking problem, addressing the optimal control problem of pursuing and escaping a non-cooperative target in space under incomplete information conditions. Zhu et al., in their paper "Optimal Control Strategy for Spacecraft Orbital Pursuit and Escape Based on Differential Game Theory," introduced the unscented Kalman filtering method into spacecraft pursuit and escape, estimating the weight parameters in the payoff function and calculating the optimal tracking control law in real time. However, both of these methods are time-consuming due to the limitations of traditional algorithms and are not efficient enough in practical applications. Summary of the Invention
[0003] The purpose of this invention is to solve the technical problem of long time consumption when using traditional algorithms to identify and invert target control strategies in the prior art, and to provide an intelligent control law inversion method for differential game strategies.
[0004] To achieve the above objectives, the present invention employs the following technical solution:
[0005] In a first aspect, the present invention provides an intelligent control law inversion method for differential game strategies, comprising:
[0006] An infinite time-domain differential game strategy model is established based on differential game strategy models and infinite time differential games;
[0007] Based on the infinite time-domain differential game strategy model, the Nash equilibrium strategy is solved to obtain the optimal strategy model of the spacecraft in the game.
[0008] Determine the input and output of the optimal strategy model, construct a dataset for the neural network, and select a suitable neural network model;
[0009] The neural network model is trained using the dataset to obtain an initial neural network model;
[0010] By adjusting the parameters and hyperparameters of the initial neural network model, the optimal intelligent control law inversion model is obtained.
[0011] A further improvement of the present invention is as follows:
[0012] The establishment of an infinite-time differential game strategy model based on differential game strategy model and infinite-time differential game specifically includes the following steps:
[0013] Assuming the pursuing spacecraft employs an infinite-time differential strategy, its cost function is:
[0014]
[0015] In the formula, Q is a positive semi-definite matrix, which is a weight matrix that measures the difference in the relative states of the two spacecraft during the game; R p and R e These are positive definite matrices, representing the weights of the energy consumption of the pursuing and escaping spacecraft in the payoff function, respectively; x(t) represents the difference in the relative states of the two spacecraft at time t; u p (t) is the control quantity of the pursuer at time t; u e (t) is the control quantity of the escapee at time t;
[0016] To obtain the saddle point solution of the infinite-time differential game, an algebraic Riccati equation is constructed:
[0017] A T P+PA+QP(B p R p -1 B p T -B e R e -1 B e T P = 0 (2)
[0018] In the formula, the Riccati matrix P is a constant matrix; A is the state matrix of the system state-space equation based on the CW linear equations; B p and B e These are the input matrices of the system state-space equations based on the CW linear equations for the pursuing and escaping spacecraft, respectively.
[0019] The process of solving the Nash equilibrium strategy based on the infinite time-domain differential game strategy model to obtain the optimal strategy model for the spacecraft in the game specifically includes the following steps:
[0020] By solving the Riccati equation, the optimal strategy model for both pursuing and fleeing spacecraft is further obtained:
[0021]
[0022] Feedback gain in the optimal strategy model As a constant matrix, this optimal strategy model is adopted when the spacecraft of the pursuing and fleeing sides engage in a game for a relatively long time.
[0023] Determining the input and output of the optimal strategy model specifically includes the following steps:
[0024] The input to the optimal strategy model is the orbital information of the pursuing spacecraft during continuous control maneuvers; the output of the optimal strategy model is the game control matrix selected by the pursuing spacecraft.
[0025] The specific steps involved in constructing the dataset for the neural network are as follows:
[0026] With a fixed weight matrix Q, different control matrices R are selected. e This generates several control strategies for pursuing spacecraft.
[0027] For different control strategies, the trajectory of the pursuing spacecraft is simulated, and a state sequence composed of the orbital position and velocity state vectors of the pursuing spacecraft at fixed time intervals is selected.
[0028] The initial position of the pursuing spacecraft is randomly set, several sets of state sequences are obtained and data calibration is performed. At the same time, the state sequences are preprocessed to complete the construction of datasets for neural network models of different control strategies.
[0029] The selection of a suitable neural network model includes:
[0030] A deep learning architecture suitable for the long short-term memory network for processing the state sequence is selected; the long short-term memory network includes an input layer, an LSTM layer, a fully connected layer, and an output layer.
[0031] The step of training the neural network model using the dataset to obtain the initial neural network model specifically includes:
[0032] The dataset is divided into a training set and a test set. The long short-term memory network is trained using the training set. The network weights are adjusted through backpropagation and gradient descent to minimize the loss function and obtain the initial neural network model.
[0033] The adjustment of network weights includes:
[0034] Given N training samples Where (n) represents the nth sample, X represents the position and velocity information of the non-cooperative target, and P represents the actual type probability distribution. For the predicted probability distribution of a type, the smaller the cross-entropy value, the closer the predicted probability distribution is to the actual type probability distribution. SoftMax classification uses cross-entropy as the loss function to learn the optimal parameter matrix ω, so the loss function is:
[0035]
[0036] The gradient of the loss function R(ω) with respect to the parameter matrix ω is:
[0037]
[0038] The training of the neural network model using the training set includes:
[0039] The gradient descent algorithm is used to iterate over the parameter ω, continuously updating ω in the direction of minimum loss along the opposite direction of the gradient; for a function R(ω), its gradient at parameter ω is... The learning rate σ is used to define the magnitude of each update, and is calculated by ω. n The gradient at the given point and the set learning rate are used to obtain the update formula:
[0040]
[0041] Compared with the prior art, the present invention has the following beneficial effects:
[0042] This invention discloses an intelligent control law inversion method for differential game strategies, proposing a linear time-invariant control matrix inversion method based on an intelligent learning structure using a neural network model. By constructing a neural network deep learning structure suitable for sequence data processing, it aims to more flexibly handle the complexity of linear time-invariant feedback control. Taking the linear feedback control law under differential games as the object, the method reveals the opponent's control strategy by inverting the unknown weight matrix and control law parameters, providing theoretical support for effectively responding to continuous thrust pursuit. Attached Figure Description
[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0044] Figure 1 This is a flowchart of an intelligent control law inversion method for differential game strategies in this invention;
[0045] Figure 2 This is a schematic diagram of an intelligent control law inversion model for differential game strategies in this invention;
[0046] Figure 3 This is a diagram showing the control law identification result based on a neural network model in an intelligent control law inversion method for differential game strategies in this invention. Figure 3 (a) is a graph of the loss function; Figure 3 (b) is the accuracy graph.
[0047] Figure 4 This is a diagram of an intelligent control law inversion system for differential game strategies in this invention;
[0048] Figure 5 This is a block diagram of the electronic device in this invention. Detailed Implementation
[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0050] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0051] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0052] The present invention will now be described in further detail with reference to the accompanying drawings:
[0053] See Figure 1 This invention discloses an intelligent control law inversion method for differential game strategies, comprising the following steps:
[0054] S1, Based on the differential game strategy model and the infinite time differential game, establish the infinite time domain differential game strategy model;
[0055] S2, based on the infinite time-domain differential game strategy model, solve the Nash equilibrium strategy to obtain the optimal strategy model of the spacecraft in the game;
[0056] S3, determine the input and output of the optimal strategy model, construct a dataset for the neural network, and select a suitable neural network model;
[0057] S4, Use the dataset to train the neural network model to obtain an initial neural network model;
[0058] S5. Adjust the parameters and hyperparameters of the initial neural network model to obtain the optimal intelligent control law inversion model.
[0059] This invention discloses an intelligent control law inversion method for differential game strategies, proposing a linear time-invariant control matrix inversion method based on an intelligent learning structure using a neural network model. By constructing a neural network deep learning structure suitable for sequence data processing, it aims to more flexibly handle the complexity of linear time-invariant feedback control. Taking the linear feedback control law under differential games as the object, the method reveals the opponent's control strategy by inverting the unknown weight matrix and control law parameters, providing theoretical support for effectively responding to continuous thrust pursuit.
[0060] See Figure 2 and Figure 3 This invention discloses an intelligent control law inversion method for differential game strategies. The following detailed description of the invention is provided in conjunction with specific embodiments:
[0061] Step 1: Establish an infinite time-domain differential game strategy model based on the differential game strategy model and the infinite time differential game;
[0062] In fixed-time differential games, if the game time between the pursuer and the pursuer is long, the game becomes an infinite-time differential game. In the infinite-time differential game, both the pursuer and the pursuer are determined to continue the game to the end and have enough fuel to burn. At this time, the payoff function changes:
[0063] Assuming this method is used to pursue the spacecraft (target), its cost function is:
[0064]
[0065] In the formula, Q is a positive semi-definite matrix, which is a weight matrix that measures the difference in the relative states of the two spacecraft during the game; R p and R e These are positive definite matrices, representing the weights of the energy consumption of the pursuing and escaping spacecraft in the payoff function, respectively; x(t) represents the difference in the relative states of the two spacecraft at time t, and u p (t) is the control variable of the pursuer at time t, u e (t) is the control quantity of the escapee at time t.
[0066] To obtain the saddle point solution of the infinite-time differential game, an algebraic Riccati equation is constructed:
[0067] A T P+PA+QP(B p R p -1 B p T -B e R e -1 B e T P = 0 (2)
[0068] In the formula, the Riccati matrix P is a constant matrix, A is the state matrix of the system state-space equation based on the CW linear equations, and B... p and B e These are the input matrices of the system state-space equations based on the CW linear equations for the pursuing and escaping spacecraft, respectively.
[0069] Step 2: Solve the Nash equilibrium strategy based on the infinite time-domain differential game strategy model to obtain the optimal strategy model of the spacecraft in the game;
[0070] By solving the Riccati equation, the optimal strategy model for both the pursuer and the pursuer is further obtained:
[0071]
[0072] Feedback gain in the above optimal strategy model The matrix is constant. When the pursuit and escape sides engage in a game for a relatively long time, this optimal strategy model is adopted.
[0073] Now assume the control matrix R of the escape spacecraft (our side) e Throughout the entire game, the pursuing spacecraft possesses complete information, namely, it knows the control matrix R of the escaping spacecraft. e .
[0074] If the pursuing spacecraft is controlled using the aforementioned infinite time-domain differential game strategy model, its control gain matrix K(t) is actually obtained by solving the algebraic Riccati equation, and depends on the weight matrices Q and R. p However, different control gain matrices result in different trajectories for the pursuing spacecraft. Therefore, the intelligent control law inversion method for differential game strategies aims to identify the control matrix adopted by the pursuing spacecraft through orbital information.
[0075] Step 3: Determine the input and output of the optimal strategy model, construct a dataset for the neural network, and select a suitable neural network model;
[0076] The orbital information of the pursuing spacecraft (target) under continuous control maneuvers is selected as the input of the optimal strategy model, and the game control matrix selected by the pursuing spacecraft (target) is selected as the output of the optimal strategy model.
[0077] With a fixed weight matrix Q, different control matrices R are selected. e This generates several control strategies for pursuing spacecraft.
[0078] For different control strategies, the trajectory of the pursuing spacecraft is simulated, and a state sequence composed of the orbital position and velocity state vectors at fixed time intervals of the pursuing spacecraft is selected.
[0079] The initial position of the pursuing spacecraft is randomly set, several sets of state sequences are obtained and data calibration is performed. At the same time, the state sequences are preprocessed to complete the construction of neural network datasets for different control strategies.
[0080] A deep learning structure suitable for processing state sequences is selected, and the entire long short-term memory network consists of an input layer, an LSTM layer, a fully connected layer, and an output layer. This structure can effectively process sequence data and achieve good performance in classification tasks.
[0081] Step 4: Train the neural network model using the dataset to obtain an initial neural network model;
[0082] The dataset is divided into a training set and a test set. The neural network model is trained using the training set. The network weights are adjusted through backpropagation and gradient descent to minimize the loss function and obtain the initial neural network model.
[0083] The research on the control law inversion method is characterized by the following process for adjusting network weights:
[0084] Given N training samples Where the superscript (n) indicates the nth sample, X represents the position and velocity information of the non-cooperative target, and P represents the actual type probability distribution. For the predicted probability distribution of a type, the smaller the cross-entropy value, the closer the predicted probability distribution is to the actual type probability distribution. SoftMax classification uses cross-entropy as the loss function to learn the optimal parameter matrix ω, so the loss function is:
[0085]
[0086] The gradient of the loss function R(ω) with respect to the parameter matrix ω is:
[0087]
[0088] Step 5: Adjust the parameters and hyperparameters of the initial neural network model to obtain the optimal intelligent control law inversion model.
[0089] Typically, the training process of deep networks can be roughly divided into two stages: the first stage is to obtain the predicted value through the forward propagation algorithm; in the second stage, the gradient of the loss function with respect to each parameter is calculated through the backpropagation algorithm, and then each parameter is updated using the gradient descent algorithm based on the gradient and the learning rate.
[0090] The gradient descent algorithm iterates over the parameter ω, continuously updating it in the opposite direction of the gradient to minimize loss. For a function R(ω), its gradient at parameter ω is... Having the gradient, we also need a learning rate σ to define the magnitude of each update, which is calculated by ω. n The update formula can be obtained by setting the gradient at a given point and the learning rate:
[0091]
[0092] For example, the control strategy is selected as follows: fixed weight matrix Q = 1 × 10 -12 I 6×6 Select different R p They are respectively 10I 3×3 and 100I 3×3 The target positions were randomly initialized for different control strategies. 30,000 sets of trajectory data were selected and calibrated, resulting in a training set of 90,000 sets. 5% of the data was randomly selected as the validation set for subsequent training. The optimal network structure and number of neurons were obtained: a 2-layer LSTM with 64 neurons using the tanh activation function; a 2-layer feedforward layer with 128 neurons using the linear activation function; and an output layer with 3 neurons using the softmax activation function, suitable for multi-class classification problems.
[0093] Weight initialization method: lecun_normal
[0094] Loss function: mean_squared_error
[0095] Optimizer: Nadam
[0096] Learning rate: 0.00001
[0097] Training results as follows Figure 3 The figure shows the results of identifying three sets of linear control matrices based on an LSTM feedforward neural network given an input sequence of observed states with a sequence length of 10. The training results show that the control matrix identification accuracy of this model on the training set can reach [percentage missing]. 99 On the validation set, the control matrix recognition accuracy of this model can reach approximately 97.5%, with an accuracy of around 0.2%. A specific embodiment is given below to illustrate the specific calculation process of this invention.
[0098] See Figure 4 This invention proposes an intelligent control law inversion system for differential game strategies, comprising:
[0099] The module for establishing infinite-time differential game strategy models establishes infinite-time differential game strategy models based on differential game strategy models and infinite-time differential games.
[0100] The optimal strategy model building module solves the Nash equilibrium strategy based on the infinite time-domain differential game strategy model to obtain the optimal strategy model of the spacecraft in the game.
[0101] The neural network model selection module determines the input and output of the optimal strategy model, constructs a dataset for the neural network, and selects a suitable neural network model.
[0102] The neural network model training module uses the dataset to train the neural network model to obtain an initial neural network model;
[0103] The intelligent control law inversion model establishment module adjusts the parameters and hyperparameters of the initial neural network model to obtain the optimal intelligent control law inversion model.
[0104] The second objective of this invention is to provide an intelligent control law inversion system for differential game strategies, proposing a linear time-invariant control matrix inversion method based on an intelligent learning structure using a neural network model. By constructing a neural network deep learning structure suitable for sequence data processing, it aims to more flexibly handle the complexity of linear time-invariant feedback control. Taking the linear feedback control law under differential games as the object, the control strategy of the opponent is revealed by inverting the unknown weight matrix and control law parameters, providing theoretical support for effectively responding to continuous thrust pursuit.
[0105] See Figure 5 A third objective of this invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the intelligent control law inversion method for differential game strategies.
[0106] An infinite time-domain differential game strategy model is established based on differential game strategy models and infinite time differential games;
[0107] Based on the infinite time-domain differential game strategy model, the Nash equilibrium strategy is solved to obtain the optimal strategy model of the spacecraft in the game.
[0108] Determine the input and output of the optimal strategy model, construct a dataset for the neural network, and select a suitable neural network model;
[0109] The neural network model is trained using the dataset to obtain an initial neural network model;
[0110] By adjusting the parameters and hyperparameters of the initial neural network model, the optimal intelligent control law inversion model is obtained.
[0111] The fourth objective of this invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the intelligent control law inversion method for differential game strategies.
[0112] An infinite time-domain differential game strategy model is established based on differential game strategy models and infinite time differential games;
[0113] Based on the infinite time-domain differential game strategy model, the Nash equilibrium strategy is solved to obtain the optimal strategy model of the spacecraft in the game.
[0114] Determine the input and output of the optimal strategy model, construct a dataset for the neural network, and select a suitable neural network model;
[0115] The neural network model is trained using the dataset to obtain an initial neural network model;
[0116] By adjusting the parameters and hyperparameters of the initial neural network model, the optimal intelligent control law inversion model is obtained.
[0117] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied 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.
[0118] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0120] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes, such as Figure 5 As shown.
[0121] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for inverting intelligent control laws for differential game strategies, characterized in that, include: An infinite time-domain differential game strategy model is established based on differential game strategy models and infinite time differential games; Based on the infinite time-domain differential game strategy model, the Nash equilibrium strategy is solved to obtain the optimal strategy model of the spacecraft in the game. Determine the input and output of the optimal strategy model, construct a dataset for the neural network, and select a suitable neural network model; Determining the input and output of the optimal strategy model specifically includes the following steps: The input to the optimal strategy model is the orbital information of the pursuing spacecraft during continuous control maneuvers; the output of the optimal strategy model is the game control matrix selected by the pursuing spacecraft. The specific steps involved in constructing the dataset for the neural network are as follows: Fixed weight matrix Selecting different control matrices This generates several control strategies for pursuing spacecraft. For different control strategies, the trajectory of the pursuing spacecraft is simulated, and a state sequence composed of the orbital position and velocity state vectors of the pursuing spacecraft at fixed time intervals is selected. The initial position of the pursuing spacecraft is randomly set, several sets of state sequences are obtained and data calibration is performed. At the same time, the state sequences are preprocessed to complete the construction of the dataset for the neural network model of different control strategies. The neural network model is trained using the dataset to obtain an initial neural network model; By adjusting the parameters and hyperparameters of the initial neural network model, the optimal intelligent control law inversion model is obtained.
2. The intelligent control law inversion method for differential game strategies according to claim 1, characterized in that, The establishment of an infinite-time differential game strategy model based on differential game strategy model and infinite-time differential game specifically includes the following steps: Assuming the pursuing spacecraft employs an infinite-time differential strategy, its cost function is: In the formula, It is a positive semi-definite matrix, which is a weight matrix that measures the difference in the relative states of the two spacecraft during the game. and These are positive definite matrices, representing the weights of the energy consumption of the pursuing and escaping spacecraft in the payoff function, respectively. express The difference in relative state between the two spacecraft at a given moment; yes The amount of control the pursuer constantly has; yes The amount of control over the escapee at any given moment; To obtain the saddle point solution of the infinite-time differential game, an algebraic Riccati equation is constructed: In the formula, the Ricardi matrix It is a constant matrix; It is the state matrix of the system state-space equation based on the CW linear equation; and These are the input matrices of the system state-space equations based on the CW linear equations for the pursuing and escaping spacecraft, respectively.
3. The intelligent control law inversion method for differential game strategies according to claim 2, characterized in that, The process of solving the Nash equilibrium strategy based on the infinite time-domain differential game strategy model to obtain the optimal strategy model for the spacecraft in the game specifically includes the following steps: By solving the Riccati equation, the optimal strategy model for both pursuing and fleeing spacecraft is further obtained: Feedback gain in the optimal strategy model As a constant matrix, this optimal strategy model is adopted when the spacecraft of the pursuing and fleeing sides engage in a game for a relatively long time.
4. The intelligent control law inversion method for differential game strategies according to claim 1, characterized in that, The selection of a suitable neural network model includes: A deep learning architecture suitable for the long short-term memory network for processing the state sequence is selected; the long short-term memory network includes an input layer, an LSTM layer, a fully connected layer, and an output layer.
5. The intelligent control law inversion method for differential game strategies according to claim 4, characterized in that, The step of training the neural network model using the dataset to obtain the initial neural network model specifically includes: The dataset is divided into a training set and a test set. The long short-term memory network is trained using the training set. The network weights are adjusted through backpropagation and gradient descent to minimize the loss function and obtain the initial neural network model.
6. The intelligent control law inversion method for differential game strategies according to claim 5, characterized in that, The adjustment of network weights includes: Given N training samples ,in, Indicates the first One sample, X For the position and velocity information of non-cooperative targets, P For the actual type probability distribution, For the predicted probability distribution of a type, the smaller the cross-entropy value, the closer the predicted probability distribution is to the actual type probability distribution. SoftMax classification uses cross-entropy as the loss function to learn the optimal parameter matrix. Then the loss function is: loss function Regarding the parameter matrix The gradient is:
7. The intelligent control law inversion method for differential game strategies according to claim 6, characterized in that, The training of the neural network model using the training set includes: Gradient descent algorithm is used to iterate parameters Continuously adjust the parameters in the opposite direction of the gradient. Update in the direction of minimizing loss; for the function Its parameters The gradient at is Learning rate Used to define the magnitude of each update, through calculation The gradient at the given point and the set learning rate are used to obtain the update formula: