Game behavior dynamic evolution and strategy deduction optimization method and system in space field

Abstract

The invention provides a space field game behavior dynamic evolution and strategy deduction optimization method and system, and relates to the technical field of space field game behavior optimization, and the method comprises the steps: analyzing historical behavior data through a deep neural network, constructing a dynamic relation graph of participants, inputting a multi-agent reinforcement learning model, and outputting a game behavior model. Then, a multi-layer game scene evolution tree is constructed, the instant income and the long-term risk are calculated in combination with a dual neural network, dynamic updating is performed by using a search algorithm of a hierarchical attention mechanism, and candidate evolution paths are generated; and finally, evaluating the candidate paths by using a strategy evaluation model constructed by a space-time diagram neural network, selecting an optimal strategy path, and generating a strategy deployment scheme implemented by stages in combination with a knowledge inference engine. According to the method, the evolution trend of the game behavior in the space field can be effectively predicted, a feasible strategy deployment scheme is provided, and the scientificity and effectiveness of strategy decision are improved.

Description

Technical Field

[0001] The present invention relates to a space field game behavior optimization technology, and in particular to a space field game behavior dynamic evolution and strategy deduction optimization method and system. Background Art

[0002] With the rapid development of space technology and the rise of commercial spaceflight, the number of participants in the space field is increasing, and their strategic game behaviors are highly complex and dynamic. Traditional game analysis methods are mostly based on static assumptions, which are difficult to effectively describe the dynamic evolution characteristics of the participants' behaviors. They simplify the decision-making behaviors of the participants into fixed strategic choices, ignoring their adaptive adjustments over time and environmental changes, resulting in a large deviation between the analysis results and the actual situation.

[0003] Existing strategy deduction methods lack systematic modeling of multi-level game scenarios, making it difficult to build a complete game scenario evolution framework and accurately grasp the overall effectiveness of the participants' decisions. Existing studies have not mined the historical behavior data of participants in depth, and lack effective data analysis models to extract the deep characteristics and evolution laws of the participants' behaviors. Strategy evaluation systems usually evaluate strategy effects from a single dimension, lacking an evaluation mechanism that comprehensively considers multiple key factors such as technical feasibility, resource constraints, and long-term benefits. In addition, traditional methods mainly rely on expert experience for strategy planning, which is difficult to cope with the rapidly changing space competition environment and cannot achieve dynamic optimization and adjustment of the strategy implementation process.

[0004] Therefore, it is urgent to propose a new method that can effectively analyze the dynamic evolution characteristics of game behavior in the space field and optimize strategies. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for the dynamic evolution and strategy deduction optimization of game behaviors in the space field, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention, A method for dynamic evolution and strategy deduction optimization of game behavior in space is provided, including: Obtain historical behavior data of participants in the space field, perform time series feature analysis on the historical behavior data through a deep neural network, obtain the periodic change law of the participants' behavior, build a feature extraction model based on the periodic change law using a distributed learning framework, generate a multi-dimensional feature vector, build a dynamic relationship graph of the participants based on the multi-dimensional feature vector, input the dynamic relationship graph into a preset multi-agent reinforcement learning model, and the multi-agent reinforcement learning model adopts an encoder structure and a multi-layer self-attention mechanism to output a game behavior model containing an adaptive utility function and dynamic decision-making rules; A multi-layer game scenario evolution tree is constructed according to the game behavior model, wherein each tree node contains the state information and decision space of the participants, the state information and decision space are input into the dual neural network structure, and the immediate benefit value and long-term risk value of the participants in each game scenario are respectively calculated. Based on the temporal variation law of the immediate benefit value and the long-term risk value, an evaluation function is constructed, and a search algorithm with a hierarchical attention mechanism is used to dynamically update the multi-layer game scenario evolution tree. The hierarchical attention mechanism adaptively adjusts the search strategy according to the evaluation function, determines the sampling weight of the evolution path, and generates multiple candidate evolution paths with different strategy characteristics; The candidate evolution path is input into the strategy evaluation model constructed by the spatiotemporal graph neural network to generate path evaluation indicators including technology maturity score, resource input score and strategy benefit score, the optimal strategy path is selected according to the path evaluation indicators, and based on the optimal strategy path, a strategy deployment plan for phased implementation is generated in combination with the knowledge reasoning engine, and the feasibility of the strategy deployment plan is evaluated and optimized through a distributed verification network.

[0007] In an optional embodiment, A dynamic relationship diagram of the participants is constructed according to the multi-dimensional feature vector, and the dynamic relationship diagram is input into a preset multi-agent reinforcement learning model. The multi-agent reinforcement learning model adopts an encoder structure and a multi-layer self-attention mechanism, and outputs a game behavior model including an adaptive utility function and dynamic decision rules, including: Generate an initial dynamic relationship graph based on the multidimensional feature vector, where nodes represent participants and graph edges represent the interaction relationship between participants; embed nodes with features through a deep neural network to obtain node representation vectors, and use a temporal convolutional network to extract interaction history to obtain dynamic pattern features; combine the node representation vector with the dynamic pattern features to calculate edge weights, and update the topological structure of the dynamic relationship graph based on the edge weights and node representation vectors; The dynamic relationship graph is input into a multi-agent reinforcement learning model, wherein the multi-agent reinforcement learning model comprises an encoder network and a strategy network; the encoder network encodes the dynamic relationship graph through a node feature layer, a relationship feature layer and a time series feature layer to obtain a state representation vector; the strategy network generates an action probability distribution of each agent based on the state representation vector; Constructing a multi-agent value network, predicting the state value and action value of each agent based on the state representation vector and the action probability distribution; inputting the state value and action value into the utility function generation module to construct a utility function including individual utility and group utility; Calculate the reward signal of each agent based on the utility function, and input the reward signal into the policy optimization module; the policy optimization module uses the policy gradient method to update the policy network parameters to achieve dynamic optimization of the agent decision rules; Adopting the experience replay mechanism to store the interaction data of the intelligent agents, and synchronously updating the value network and utility function parameters based on the interaction data; realizing the joint optimization of the strategy network, value network and utility function through multi-agent collaborative learning; The optimized strategy network, value network and utility function are integrated into a model to output a game behavior model that includes an adaptive utility function and dynamic decision rules.

[0008] In an optional embodiment, Constructing a multi-agent value network, predicting the state value and action value of each agent based on the state representation vector and action probability distribution; inputting the state value and action value into the utility function generation module, and constructing a utility function containing individual utility and group utility includes: Construct a state value network and an action value network. The state value network adopts a three-layer fully connected network structure, and the state representation vector is input to obtain the state value prediction value; the action value network adopts a double-branch structure, the first branch processes the state representation vector, and the second branch processes the action probability distribution. The two branch features are fused through the outer product operation and then the action value prediction value is obtained through the fully connected layer; Calculate a time series difference target value, the time series difference target value is calculated based on the immediate reward value, the state value prediction value and the action value prediction value; construct a state value loss function by the difference between the time series difference target value and the state value prediction value, and construct an action value loss function by the difference between the time series difference target value and the action value prediction value; update the value network parameters based on the state value loss function and the action value loss function to obtain optimized state value and action value; The optimized action value is used to represent the individual utility term, and the group utility term is obtained by weighted calculation of the optimized state value and the action value of the neighborhood agent; an initial utility function is constructed by combining the individual utility term and the group utility term; The gradient value of the initial utility function with respect to the weight coefficient is calculated, and based on the gradient value, the weight coefficients of the individual utility term and the group utility term are updated using the gradient descent method to obtain the final utility function.

[0009] In an optional embodiment, A multi-layer game scenario evolution tree is constructed according to the game behavior model, wherein each tree node contains the state information and decision space of the participants, and the state information and decision space are input into the dual neural network structure to calculate the immediate benefit value and long-term risk value of the participants in each game scenario respectively, including: A multi-layer game scenario evolution tree is constructed based on the dynamic decision rules in the game behavior model, and the state vectors and decision space information of the participants are stored in the tree nodes; the transition probability matrix between adjacent nodes is calculated based on the state transition function; Monte Carlo sampling is performed according to the transition probability matrix to generate an evolution path from the root node to the leaf node; the node sequence corresponding to the evolution path and the historical state sequence of the participants are stored as feature inputs in each tree node; Construct a dual neural network structure, including a graph attention network and a hierarchical recurrent neural network, input the state vector in the tree node into the graph attention network, generate a feature representation matrix of the participants through the feature transformation layer, calculate the attention coefficient between the participants based on the feature representation matrix, multiply the attention coefficient with the feature representation matrix to obtain a weighted feature, and perform a residual connection between the weighted feature and the feature representation matrix, transform the residually connected features through a multi-head attention mechanism, extract the correlation features between the participants, and input the correlation features into the adaptive pooling layer of the graph attention network for dimensionality reduction aggregation to obtain an instantaneous benefit value reflecting the game state of the participants at the current node; The historical state sequences stored in the tree nodes are grouped according to the participant identifiers, and the grouped sequences are input into a hierarchical recurrent neural network. The recurrent neural network selectively memorizes the grouped sequences through a gating unit, outputs the state features of each time step, calculates the attention weight of the time dimension for the state features, and obtains a temporal correlation representation that describes the state evolution law. At the same time, the attention weight of the participant dimension is calculated for the state features to obtain an interactive correlation representation that describes the game relationship. The temporal correlation representation and the interactive correlation representation are weightedly combined with the state features, and the long-term risk value reflecting the cumulative effect of the participants on the evolution path is output through the prediction module of the hierarchical recurrent neural network.

[0010] In an optional embodiment, Based on the temporal variation law of the instant return value and the long-term risk value, an evaluation function is constructed, and a search algorithm with a hierarchical attention mechanism is used to dynamically update the multi-layer game scenario evolution tree. The hierarchical attention mechanism adaptively adjusts the search strategy according to the evaluation function, determines the sampling weight of the evolution path, and generates multiple candidate evolution paths with different strategy characteristics, including: Calculate the rate of change of benefits according to the instant benefit value of the game state of adjacent nodes, and calculate the risk change trend according to the long-term risk value of the historical state sequence; construct a time series evaluation function by weighted combination of the instant benefit value, the rate of change of benefits and the risk change trend; calculate the time series evaluation score of each tree node in the multi-layer game scenario evolution tree according to the time series evaluation function; A state attention layer is constructed based on the temporal evaluation scores of tree nodes. The state importance weights of tree nodes are output through the state attention layer. At the same time, the temporal evaluation scores and corresponding state importance weights of adjacent tree nodes are input into the transfer attention layer to obtain the importance weights of state transfers. The complete path sampling probability from the root node to the leaf node is calculated based on the state importance weights of the tree nodes and the importance weights of state transfers. The temperature parameter is introduced to adjust the path sampling probability, and the gradient change value of the timing evaluation function is detected. When the gradient change value is less than the convergence threshold, the temperature parameter is increased, and the timing evaluation score of the sampling path is detected. When the score exceeds the historical optimal, the temperature parameter is reduced. The ratio of the path sampling probability to the temperature parameter is used as the adjusted sampling probability, and the candidate evolution path is generated based on the adjusted sampling probability. Maintain a fixed-size candidate path set, calculate the weighted sum of the temporal evaluation scores of all tree nodes on the newly generated path as the total evaluation score of the path, and when the total evaluation score of the new path is higher than the lowest score in the set, add the new path to the candidate path set and remove the path with the lowest score; The edit distance between paths in the candidate path set is calculated to obtain the path diversity index, and the temperature parameter is dynamically adjusted based on the diversity index. When the diversity index is lower than the target threshold, the temperature parameter is increased to expand the search range. When the diversity index is higher than the target threshold, the temperature parameter is reduced to strengthen the local search. The path weight is calculated according to the total evaluation score and sampling probability of each candidate path, and the multi-layer game scenario evolution tree is updated based on the path weight, and finally multiple candidate evolution paths with different strategic characteristics are obtained.

[0011] In an optional embodiment, The candidate evolution path is input into the strategy evaluation model constructed by the spatiotemporal graph neural network to generate a path evaluation index including a technology maturity score, a resource input score and a strategy benefit score. The optimal strategy path is selected according to the path evaluation index, including: Constructing a spatiotemporal heterogeneous graph structure, taking technology nodes and resource nodes as heterogeneous graph nodes, setting the technology development sequence relationship as a temporal edge, setting the resource dependency relationship as a spatial edge, constructing a technology node attribute vector based on technology attribute characteristics and timestamp information, constructing a resource node attribute vector based on resource type and capacity information, and combining the technology node attribute vector and resource node attribute vector to form a node feature matrix; The node feature matrix and the temporal edge are input into the causal convolutional network to extract the temporal context features, and the feature weights are calculated through the temporal attention mechanism to obtain enhanced temporal features; the node feature matrix and the spatial edge are input into the graph attention network to extract the node interaction features, and the feature importance is calculated through the spatial attention mechanism to obtain enhanced spatial features; the temporal features and spatial features are input into the gated update unit, the gated weights are calculated, and adaptive fusion is performed to obtain a unified feature representation; Based on the unified feature representation, a multi-layer perceptron and an expert knowledge base are used to perform feature matching to obtain a technical feasibility score; a graph matching network is used to calculate the matching degree between the path resource requirements and the template resources to obtain a resource consumption score; and a graph attention network is used to extract strategy-related features to obtain a strategy value score; Based on the unified feature representation, the mixed Gaussian distribution parameters are predicted, and in combination with the historical evaluation data, the posterior distribution is calculated by variational inference to perform Bayesian calibration on the technical feasibility score, resource consumption score, and strategic value score to obtain a calibrated score; The calibrated score is input into the path optimization objective function, and iterative optimization is performed through a beam search algorithm under the constraints of technical feasibility, resource consumption and strategic value, and finally the optimal strategic path is obtained.

[0012] In an optional embodiment, Based on the optimal strategy path, a strategy deployment plan implemented in stages is generated in combination with a knowledge reasoning engine, and the feasibility of the strategy deployment plan is evaluated and optimized through a distributed verification network, including: Constructing a policy deployment knowledge base, the policy deployment knowledge base including a rule base of technology dependency rules, resource allocation rules and progress control rules, and a fact base of acquired technology lists, current available resource status and historical deployment case data; A knowledge reasoning engine is constructed based on the rule base and the fact base. The knowledge reasoning engine extracts the technical dependency relationship of each technical node in the optimal strategy path to obtain a dependency chain. The earliest implementable time of each technical node is calculated by combining the dependency chain and the current system state. The earliest implementable time is matched with the progress control rule in the rule base to generate a time execution window. A resource scheduling plan is calculated based on the time execution window and the resource allocation rule to form a phased deployment plan. Construct a hierarchical verification architecture, which includes a first-level verification component that performs global consistency checks, a second-level verification component that performs technology dependency verification, and a third-level verification component that performs resource constraint verification, and distributes verification tasks to verification components at each level according to a hierarchical authorization mechanism; The verification results of verification components at all levels are input into the Byzantine Fault Tolerant consensus algorithm, and the credibility score of each verification result is calculated based on the reputation weight. The verification results are sorted according to the credibility score, and the highly credible verification opinions are structured to obtain optimization suggestions. Construct a verification indicator vector based on the optimization suggestions, input the verification data of technical feasibility, resource matching and schedule controllability into the corresponding indicator calculation model, and generate a quantitative verification evaluation result; The verification and evaluation results are input into the policy optimization network to generate targeted optimization and adjustment plans, which are applied to the execution plan for updating. The updated execution plan is re-input into the hierarchical verification architecture for evaluation. When the verification and evaluation results meet the preset standards, the final policy deployment plan is output. When the verification and evaluation results do not meet the standards, the execution plan is returned to the policy optimization network for iterative optimization.

[0013] A second aspect of an embodiment of the present invention provides a system for dynamic evolution and strategy deduction optimization of game behavior in space, including: The first unit is used to obtain historical behavior data of participants in the space field, perform time series feature analysis on the historical behavior data through a deep neural network, obtain the periodic change law of the participants' behavior, and build a feature extraction model based on the periodic change law using a distributed learning framework to generate a multi-dimensional feature vector. A dynamic relationship diagram of the participants is built according to the multi-dimensional feature vector, and the dynamic relationship diagram is input into a preset multi-agent reinforcement learning model. The multi-agent reinforcement learning model adopts an encoder structure and a multi-layer self-attention mechanism to output a game behavior model including an adaptive utility function and dynamic decision-making rules; The second unit is used to construct a multi-layer game scenario evolution tree according to the game behavior model, wherein each tree node contains the state information and decision space of the participants, input the state information and decision space into the dual neural network structure, respectively calculate the immediate benefit value and long-term risk value of the participants in each game scenario, and construct an evaluation function based on the temporal variation law of the immediate benefit value and the long-term risk value. The multi-layer game scenario evolution tree is dynamically updated by using a search algorithm with a hierarchical attention mechanism. The hierarchical attention mechanism adaptively adjusts the search strategy according to the evaluation function, determines the sampling weight of the evolution path, and generates multiple candidate evolution paths with different strategy characteristics; The third unit is used to input the candidate evolution path into the strategy evaluation model constructed by the spatiotemporal graph neural network, generate path evaluation indicators including technology maturity score, resource input score and strategy benefit score, select the optimal strategy path according to the path evaluation indicators, and generate a strategy deployment plan for phased implementation based on the optimal strategy path in combination with the knowledge reasoning engine, and evaluate and optimize the feasibility of the strategy deployment plan through a distributed verification network.

[0014] A third aspect of the embodiments of the present invention An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0016] In this embodiment, the time series feature analysis of historical behavior data is performed through a deep neural network, and combined with a multi-agent reinforcement learning model, it is possible to more accurately predict the behavior of participants in the space field, reveal their periodic change patterns and dynamic relationships, and build a game behavior model that is closer to reality. Based on the game behavior model, a multi-layer game scenario evolution tree is constructed, and the search algorithm of the dual neural network and hierarchical attention mechanism is combined to dynamically evaluate the immediate benefits and long-term risks of different strategies, generate multiple candidate evolution paths with different strategy characteristics, and select the optimal strategy path from them. The strategy evaluation model is constructed using a spatiotemporal graph neural network, and combined with a knowledge reasoning engine to generate a phased strategy deployment plan, which can evaluate and optimize the feasibility of the strategy deployment plan and improve the efficiency and success rate of strategy deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the method for dynamic evolution and strategy deduction optimization of game behavior in the space field according to an embodiment of the present invention; Figure 2 It is a structural diagram of the system for dynamic evolution and strategy deduction and optimization of game behaviors in the space field according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0020] Figure 1 Schematic diagram of the process of the dynamic evolution and strategy deduction optimization method of the space field game behavior in an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Obtain historical behavior data of participants in the space field, perform time series feature analysis on the historical behavior data through a deep neural network, obtain the periodic change law of the participants' behavior, build a feature extraction model based on the periodic change law using a distributed learning framework, generate a multi-dimensional feature vector, build a dynamic relationship graph of the participants based on the multi-dimensional feature vector, input the dynamic relationship graph into a preset multi-agent reinforcement learning model, the multi-agent reinforcement learning model adopts an encoder structure and a multi-layer self-attention mechanism, and outputs a game behavior model including an adaptive utility function and dynamic decision-making rules; S102. Construct a multi-layer game scenario evolution tree according to the game behavior model, wherein each tree node contains the state information and decision space of the participants, input the state information and decision space into the dual neural network structure, respectively calculate the immediate benefit value and long-term risk value of the participants in each game scenario, construct an evaluation function based on the temporal variation law of the immediate benefit value and the long-term risk value, and dynamically update the multi-layer game scenario evolution tree using a search algorithm with a hierarchical attention mechanism, wherein the hierarchical attention mechanism adaptively adjusts the search strategy according to the evaluation function, determines the sampling weight of the evolution path, and generates multiple candidate evolution paths with different strategy characteristics; S103. Input the candidate evolution path into the strategy evaluation model constructed by the spatiotemporal graph neural network to generate path evaluation indicators including technology maturity score, resource input score and strategy benefit score, select the optimal strategy path according to the path evaluation indicators, and based on the optimal strategy path, generate a strategy deployment plan for phased implementation in combination with the knowledge reasoning engine, and evaluate and optimize the feasibility of the strategy deployment plan through a distributed verification network.

[0021] Among them, historical behavior data refers to the decision-making records, resource allocation and behavior patterns of the participants in the past period of time. Through time series feature analysis by deep neural networks, the periodic change patterns of the behaviors of the participants can be extracted. Time series feature analysis can identify periodic fluctuations and trend changes in behavior patterns, such as possible strategy adjustments or resource allocation changes of participants in a specific period.

[0022] Based on the extracted periodic change rules, a feature extraction model is constructed using a distributed learning framework. The distributed learning framework can efficiently process large-scale data sets and map multidimensional data to feature space. The multidimensional feature vectors generated by this process will be used to describe the behavioral characteristics and decision-making environment of the participants. Through the multidimensional feature vectors, a dynamic relationship diagram of the participants can be constructed, which can reflect the mutual influence and behavioral relationship between different participants.

[0023] In an optional implementation, a dynamic relationship diagram of the participants is constructed according to the multi-dimensional feature vector, and the dynamic relationship diagram is input into a preset multi-agent reinforcement learning model. The multi-agent reinforcement learning model adopts an encoder structure and a multi-layer self-attention mechanism, and outputs a game behavior model including an adaptive utility function and dynamic decision rules, including: Generate an initial dynamic relationship graph based on the multidimensional feature vector, where nodes represent participants and graph edges represent the interaction relationship between participants; embed nodes with features through a deep neural network to obtain node representation vectors, and use a temporal convolutional network to extract interaction history to obtain dynamic pattern features; combine the node representation vector with the dynamic pattern features to calculate edge weights, and update the topological structure of the dynamic relationship graph based on the edge weights and node representation vectors; The dynamic relationship graph is input into a multi-agent reinforcement learning model, wherein the multi-agent reinforcement learning model comprises an encoder network and a strategy network; the encoder network encodes the dynamic relationship graph through a node feature layer, a relationship feature layer and a time series feature layer to obtain a state representation vector; the strategy network generates an action probability distribution of each agent based on the state representation vector; Constructing a multi-agent value network, predicting the state value and action value of each agent based on the state representation vector and the action probability distribution; inputting the state value and action value into the utility function generation module, and constructing a utility function including individual utility and group utility; Calculating a reward signal for each agent based on the utility function, and inputting the reward signal into a policy optimization module; the policy optimization module uses a policy gradient method to update policy network parameters to achieve dynamic optimization of the agent's decision rules; Adopting the experience replay mechanism to store the interaction data of the intelligent agents, and synchronously updating the value network and utility function parameters based on the interaction data; realizing the joint optimization of the strategy network, value network and utility function through multi-agent collaborative learning; The optimized strategy network, value network and utility function are integrated into a model to output a game behavior model that includes an adaptive utility function and dynamic decision rules.

[0024] Exemplarily, first, an initial dynamic relationship graph is generated through a multidimensional feature vector. The nodes of the graph represent the various participants in the space field, and the edges in the graph represent the interactive relationships between these participants, such as cooperation, competition, resource sharing, etc. When generating the initial dynamic relationship graph, the weight of the edge can be set according to factors such as the interaction intensity and cooperation frequency between the participants. A deep neural network is used to embed features of each node. This process converts the attributes, behavior patterns, strategy selection and other information of each participant into a vector, called a node representation vector. Then, a temporal convolutional network (TCN) is used to analyze the interaction history and extract dynamic pattern features to help the system identify the behavioral changes of the participants at different time points, such as periodic resource allocation, strategy adjustment, etc. These dynamic pattern features are combined with the node representation vector, and the topological structure of the dynamic relationship graph is adjusted by calculating the weights of the graph edges, thereby reflecting the changing relationship between the participants.

[0025] After the dynamic relationship graph is constructed, it is input into the multi-agent reinforcement learning model, which includes an encoder network and a policy network. The role of the encoder network is to convert the complex information in the dynamic relationship graph into a state representation vector, which will serve as the basis for the agent's decision-making. The encoder network includes multiple feature layers, each of which processes different types of information. Specifically, the node feature layer is responsible for encoding the attributes and behavior information of the nodes, the relationship feature layer encodes the interaction between the participants, and the time series feature layer specifically extracts the pattern of the participants' behavior changing over time. Through the combination of these layers, the encoder network is able to generate a comprehensive state representation vector that represents the status of each participant in the current game scenario.

[0026] Next, the policy network generates the action probability distribution of each agent based on this state representation vector. In the game environment, each agent represents a participant, and the action probability distribution represents the probability of different strategies that the participant may take in the current state. The policy network will generate a strategy distribution for each participant based on the input state information to determine their behavior in the current game scenario.

[0027] In order to evaluate the value of each agent taking different strategies in a specific state, a multi-agent value network is constructed. The role of this network is to predict the state value and action value of each agent based on the state representation vector and action probability distribution. The state value represents the overall value of the environment in which the agent is in a certain state, while the action value is the possible reward of the agent taking a certain action in this state. By calculating the state value and action value, the value network helps the system evaluate the impact of different strategies on the game process.

[0028] On this basis, a utility function generation module is constructed. The utility function consists of two parts: individual utility and group utility. Individual utility reflects the short-term and long-term interests of each participant in the game, while group utility evaluates the cooperation and competition effects of all participants. The utility function combines state value, action value and other environmental factors to provide a reward signal for the agent, which will be used to optimize the agent's decision-making strategy.

[0029] The reward signal is input into the policy optimization module and optimized using the policy gradient method. The policy gradient method updates the parameters of the policy network so that the action selection of each agent in different states is more in line with the optimal strategy. As the policy network is continuously optimized, the decision rules of the agent will be gradually adjusted dynamically, so that the game behavior model can more accurately reflect the strategic choices of the participants in complex games.

[0030] In order to further improve the learning effect, the experience replay mechanism is adopted. The experience replay mechanism allows the model to review and utilize previous experience in the subsequent learning process by storing the interaction data of the agents. This can prevent the model from falling into the local optimal solution and accelerate the process of multi-agent collaborative learning. Based on these interaction data, the model will synchronously update the parameters of the value network and utility function, thereby achieving parameter optimization in continuous feedback.

[0031] The key to multi-agent collaborative learning is to jointly optimize the strategy network, value network and utility function through the interaction and information sharing between multiple agents. This collaborative learning mechanism ensures that the system can consider the needs and interests of different participants from a global perspective and optimize the overall game behavior.

[0032] Finally, through model integration, the optimized strategy network, value network and utility function are integrated to output a game behavior model that includes an adaptive utility function and dynamic decision-making rules. This model can not only predict the behavior of the participants in the game process, but also adaptively adjust the strategy according to the changes in the game environment, thus providing a scientific basis and decision-making support for multi-party games and strategic decisions in the space field.

[0033] In this embodiment, the dynamic behavior and strategic decision-making of multiple participants in the space field are accurately modeled through deep learning and reinforcement learning technology. Through historical behavior data analysis and multi-dimensional feature vector extraction, the system can capture the complex interactive relationships and behavior patterns between the participants, and reflect their real-time changes through dynamic relationship diagrams. The multi-agent reinforcement learning model effectively predicts the behavior of all parties and optimizes the decision-making process through adaptive decision rules and game behavior modeling, realizing collaborative learning and dynamic optimization in the game process. Through the utility function and strategy optimization module, the system can not only improve the benefits of individual participants, but also optimize the overall benefits of the group. The experience replay mechanism ensures that the model can learn efficiently and avoid local optimality, and promotes the linkage optimization of strategies, value networks and utility functions. This technical solution can provide scientific decision-making basis for complex multi-party games in the space field, support multi-party strategy coordination and long-term planning optimization, improve decision-making efficiency and accuracy, and has broad application prospects.

[0034] In an optional implementation, a multi-agent value network is constructed, and the state value and action value of each agent are predicted based on the state representation vector and the action probability distribution; the state value and action value are input into the utility function generation module, and the utility function including individual utility and group utility is constructed, including: Construct a state value network and an action value network. The state value network adopts a three-layer fully connected network structure, and the state representation vector is input to obtain the state value prediction value; the action value network adopts a double-branch structure, the first branch processes the state representation vector, and the second branch processes the action probability distribution. The two branch features are fused through the outer product operation and then the action value prediction value is obtained through the fully connected layer; Calculate a time series difference target value, the time series difference target value is calculated based on the immediate reward value, the state value prediction value and the action value prediction value; construct a state value loss function by the difference between the time series difference target value and the state value prediction value, and construct an action value loss function by the difference between the time series difference target value and the action value prediction value; update the value network parameters based on the state value loss function and the action value loss function to obtain optimized state value and action value; The optimized action value is used to represent the individual utility term, and the group utility term is obtained by weighted calculation of the optimized state value and the action value of the neighborhood agent; an initial utility function is constructed by combining the individual utility term and the group utility term; The gradient value of the initial utility function with respect to the weight coefficient is calculated, and based on the gradient value, the weight coefficients of the individual utility term and the group utility term are updated using the gradient descent method to obtain the final utility function.

[0035] Exemplarily, first, prepare a multi-agent environment and interaction data between agents. Collect the reward values ​​obtained by each agent after performing various actions in different states, as well as the state transition information of the environment. Next, construct a state value network and an action value network. The state value network is used to predict the long-term value of an agent in a specific state. It adopts a three-layer fully connected network structure, the input is a vector representation of the state, such as the position, speed, surrounding environment information of the agent, etc., and the output is a numerical value representing the value of the state. The action value network is used to predict the long-term value of an agent performing a specific action in a specific state. It adopts a dual-branch structure, one branch processes the state vector and the other branch processes the action probability distribution. The outputs of the two branches are fused through an outer product operation, and then pass through a fully connected layer to finally output a numerical value representing the value of performing the action in this state. Assuming that the state vector dimension is 5 and the action probability distribution dimension is 3, the feature dimension obtained after the outer product operation is 15, and then passes through a fully connected layer with an output dimension of 1 to obtain the action value prediction value.

[0036] Then, the TD target is calculated. This value is used to evaluate whether the current value prediction is accurate. It combines the current reward, the value of the next state, and the value of the action performed in the current state. For example, if the current reward is 10, the value prediction of the next state is 20, and the value prediction of the action performed in the current state is 15, then the TD target can be calculated as 10+0.9*20-15=13, where 0.9 is a discount factor used to balance the importance of current rewards and future rewards.

[0037] Next, construct and minimize the loss function. Calculate the difference between the time series difference target value and the state value prediction value, and the difference between the time series difference target value and the action value prediction value. These two differences constitute the state value loss function and the action value loss function respectively. By minimizing these two loss functions, the parameters of the value network can be updated, thereby improving the accuracy of value prediction.

[0038] After that, construct the utility function. The utility function is used to measure the value obtained by the agent. It contains individual utility terms and group utility terms. The individual utility term is represented by the optimized action value, and the group utility term is calculated by weighting the state value of the current agent and the action value of its neighboring agents. For example, if an individual utility value is 15, the action values ​​of neighboring agents are 10 and 12, and the weights are 0.6 and 0.4, respectively, then the group utility term is 0.6*10+0.4*12=10.8. The initial utility function can be defined as a linear combination of individual utility and group utility terms, such as 0.8*individual utility+0.2*group utility.

[0039] Finally, optimize the utility function. Calculate the gradient value of the initial utility function with respect to the weight coefficient, and then use the gradient descent method to update the weight coefficients of the individual utility term and the group utility term. For example, if the initial value of the weight coefficient of the individual utility term is 0.8, the gradient value is -0.1, and the learning rate is 0.01, then the updated weight coefficient is 0.8-0.01*(-0.1)=0.801. By continuously iteratively updating the weight coefficient, the final utility function can be obtained.

[0040] In this embodiment, by considering individual utility and group utility, agents are encouraged to promote the realization of overall goals while pursuing their own interests, thereby improving the collaborative efficiency of the multi-agent system. By constructing a state value network and an action value network, and combining the temporal difference learning method, the long-term value of agents performing different actions in different states can be more accurately evaluated. By optimizing the weight coefficient of the utility function through the gradient descent method, the method can be adapted to different multi-agent environments and tasks, and has stronger generalization ability.

[0041] In an optional implementation, a multi-layer game scenario evolution tree is constructed according to the game behavior model, wherein each tree node contains the state information and decision space of the participant, and the state information and decision space are input into a dual neural network structure, and the immediate benefit value and long-term risk value of the participant in each game scenario are calculated respectively, including: A multi-layer game scenario evolution tree is constructed based on the dynamic decision rules in the game behavior model, and the state vectors and decision space information of the participants are stored in the tree nodes; the transition probability matrix between adjacent nodes is calculated based on the state transition function; Monte Carlo sampling is performed according to the transition probability matrix to generate an evolution path from the root node to the leaf node; the node sequence corresponding to the evolution path and the historical state sequence of the participants are stored as feature inputs in each tree node; Construct a dual neural network structure, including a graph attention network and a hierarchical recurrent neural network, input the state vector in the tree node into the graph attention network, generate a feature representation matrix of the participants through the feature transformation layer, calculate the attention coefficient between the participants based on the feature representation matrix, multiply the attention coefficient with the feature representation matrix to obtain a weighted feature, and perform a residual connection between the weighted feature and the feature representation matrix, transform the residually connected features through a multi-head attention mechanism, extract the correlation features between the participants, and input the correlation features into the adaptive pooling layer of the graph attention network for dimensionality reduction aggregation to obtain an instantaneous benefit value reflecting the game state of the participants at the current node; The historical state sequences stored in the tree nodes are grouped according to the participant identifiers, and the grouped sequences are input into a hierarchical recurrent neural network. The recurrent neural network selectively memorizes the grouped sequences through a gating unit, outputs the state features of each time step, calculates the attention weight of the time dimension for the state features, and obtains a temporal correlation representation that describes the state evolution law. At the same time, the attention weight of the participant dimension is calculated for the state features to obtain an interactive correlation representation that describes the game relationship. The temporal correlation representation and the interactive correlation representation are weightedly combined with the state features, and the long-term risk value reflecting the cumulative effect of the participants on the evolution path is output through the prediction module of the hierarchical recurrent neural network.

[0042] Exemplarily, first, a multi-layer game scenario evolution tree is constructed based on a predefined game behavior model. The game behavior model describes the possible decisions of the participants in different states and their probability distribution. Each node of the evolution tree represents a specific game scenario, containing the state information of the participants, such as resource ownership, reputation value, etc., and the decision space in the scenario, such as cooperation, competition, exit, etc. The root node of the tree represents the initial game state, and the child nodes represent the subsequent possible game scenarios. The connection between the parent node and the child node represents the state transfer, which is determined by the state transfer function. The state transfer function determines the probability of transferring from the current state to the next state based on the decisions of the participants and environmental factors. For example, if both parties choose to cooperate, the probability of transferring to a state where both parties' benefits increase is higher; if one party chooses to compete, the probability of transferring to a state where the benefits of the party increase but the benefits of the other party decrease is higher.

[0043] Next, calculate the transition probability matrix between adjacent nodes. Each element of the transition probability matrix represents the probability of transitioning from one node to another. By analyzing the decision space and state transition function of the participants, the probability of each state transition can be calculated. For example, suppose that at a certain node, participant A has three decision options: cooperation, competition, and exit, and participant B also has three decision options: cooperation, competition, and exit. Then, the number of child nodes of this node is at most nine, representing all possible decision combinations. According to the game behavior model, the probability of each decision combination leading to state transition can be determined, thereby constructing a transition probability matrix.

[0044] Then, the Monte Carlo sampling method is used to generate an evolutionary path from the root node to the leaf node. Monte Carlo sampling is a random simulation method that can simulate the evolution of the game scenario based on the transition probability matrix. Specifically, starting from the root node, a child node is randomly selected as the next state according to the transition probability, and this process is repeated until a leaf node is reached. Repeating Monte Carlo sampling multiple times can generate multiple evolutionary paths, representing multiple possibilities of the game process. The node sequences corresponding to these evolutionary paths and the historical state sequences of the participants are stored in the corresponding tree nodes. For example, an evolutionary path may contain three nodes, representing three consecutive game scenarios, each of which stores the state information and decision space of the participants in the scenario, as well as the sequence of participant state changes from the root node to the node.

[0045] Subsequently, a dual neural network structure is constructed to evaluate the benefits and risks of the participants. The dual neural network structure includes a graph attention network and a hierarchical recurrent neural network. First, the state vector in the tree node is input into the graph attention network. The graph attention network converts the state vector into a feature representation matrix through a feature transformation layer. Then, the attention coefficient between the participants is calculated based on the feature representation matrix. The attention coefficient reflects the degree of mutual influence between the participants. The attention coefficient is multiplied by the feature representation matrix to obtain the weighted feature, and the weighted feature is residually connected to the feature representation matrix, and then transformed through the multi-head attention mechanism to extract the correlation features between the participants. Finally, the correlation features are input into the adaptive pooling layer of the graph attention network for dimensionality reduction aggregation to obtain the instantaneous benefit value reflecting the game state of the participants at the current node.

[0046] At the same time, the historical state sequences stored in the tree nodes are grouped according to the participant identifiers, and the grouped sequences are input into the hierarchical recurrent neural network. The recurrent neural network selectively memorizes the grouped sequences through the gated unit and outputs the state features of each time step. The attention weight of the time dimension is calculated for the state features to obtain the temporal correlation representation that describes the state evolution law. At the same time, the attention weight of the participant dimension is calculated for the state features to obtain the interactive correlation representation that describes the game relationship. The temporal correlation representation and the interactive correlation representation are weightedly combined with the state features, and the prediction module of the hierarchical recurrent neural network outputs the long-term risk value that reflects the cumulative effect of the participants on the evolution path.

[0047] For example, suppose there are two participants A and B, and the state vector at a certain node is the amount of resources of A and the amount of resources of B. The graph attention network can learn the relationship between the amounts of resources of A and B, such as a competitive relationship or a cooperative relationship, and output the immediate benefit value of A and B at this node. At the same time, the hierarchical recurrent neural network can learn the evolutionary laws and interactive relationships in the historical state sequence of A and B, for example, an increase in the amount of resources of A will lead to a decrease in the amount of resources of B, and output the long-term risk value of A and B on this evolutionary path.

[0048] In this embodiment, by combining the graph attention network and the hierarchical recurrent neural network, the mutual influence and state evolution law between the participants can be considered more comprehensively, so as to more accurately evaluate the benefits and risks of the participants. By generating multiple evolution paths through Monte Carlo sampling, multiple possibilities in the game process are considered, thereby improving the reliability of game prediction. This method can provide more sophisticated game analysis results, help participants better understand the game situation, and formulate more effective game strategies.

[0049] In an optional implementation, based on the temporal variation law of the instant benefit value and the long-term risk value, an evaluation function is constructed, and a search algorithm with a hierarchical attention mechanism is used to dynamically update the multi-layer game scenario evolution tree. The hierarchical attention mechanism adaptively adjusts the search strategy according to the evaluation function, determines the sampling weight of the evolution path, and generates multiple candidate evolution paths with different strategy characteristics, including: Calculate the rate of change of benefits according to the instant benefit value of the game state of adjacent nodes, and calculate the risk change trend according to the long-term risk value of the historical state sequence; construct a time series evaluation function by weighted combination of the instant benefit value, the rate of change of benefits and the risk change trend; calculate the time series evaluation score of each tree node in the multi-layer game scenario evolution tree according to the time series evaluation function; A state attention layer is constructed based on the temporal evaluation scores of tree nodes. The state importance weights of tree nodes are output through the state attention layer. At the same time, the temporal evaluation scores and corresponding state importance weights of adjacent tree nodes are input into the transfer attention layer to obtain the importance weights of state transfers. The complete path sampling probability from the root node to the leaf node is calculated based on the state importance weights of the tree nodes and the importance weights of state transfers. The temperature parameter is introduced to adjust the path sampling probability, and the gradient change value of the timing evaluation function is detected. When the gradient change value is less than the convergence threshold, the temperature parameter is increased, and the timing evaluation score of the sampling path is detected. When the score exceeds the historical optimal, the temperature parameter is reduced. The ratio of the path sampling probability to the temperature parameter is used as the adjusted sampling probability, and the candidate evolution path is generated based on the adjusted sampling probability. Maintain a fixed-size candidate path set, calculate the weighted sum of the temporal evaluation scores of all tree nodes on the newly generated path as the total evaluation score of the path, and when the total evaluation score of the new path is higher than the lowest score in the set, add the new path to the candidate path set and remove the path with the lowest score; The edit distance between paths in the candidate path set is calculated to obtain the path diversity index, and the temperature parameter is dynamically adjusted based on the diversity index. When the diversity index is lower than the target threshold, the temperature parameter is increased to expand the search range. When the diversity index is higher than the target threshold, the temperature parameter is reduced to strengthen the local search. The path weight is calculated according to the total evaluation score and sampling probability of each candidate path, and the multi-layer game scenario evolution tree is updated based on the path weight, and finally multiple candidate evolution paths with different strategic characteristics are obtained.

[0050] For example, first, the system analyzes the adjacent nodes in the multi-layer game scenario evolution tree and calculates the rate of change of benefits and the trend of risk changes between each node. The rate of change of benefits is reflected by the difference in the immediate benefit values ​​between adjacent nodes, indicating the growth rate of the participants' returns at different decision-making stages; the trend of risk changes is based on the long-term risk value of the adjacent nodes, assessing the long-term uncertainty and risks that the participants may face in the game. For example, when the strategic choices of the participants change greatly, it may lead to rapid fluctuations in benefits or sharp changes in risks. The system can identify potential decision-making bottlenecks and turning points through these changing patterns.

[0051] Based on the above analysis, the system weights the instantaneous return value, the rate of change of return, and the trend of risk change to construct a time series evaluation function. This evaluation function reflects the evolution of the game scenario on the timeline and can comprehensively evaluate the decision-making process of the participants and their impact on the overall game. In the multi-layer game scenario evolution tree, each tree node represents a game state. The time series evaluation function calculates the time series evaluation score of each node based on the node's instantaneous return, rate of change of return, and trend of risk change.

[0052] Next, the system built a state attention layer and a transfer attention layer. The state attention layer is used to calculate the state importance weights of tree nodes, and by analyzing the timing evaluation scores, it identifies the importance of the current game state to the global game. For example, at some decision nodes, the strategy adjustments of some participants may have a decisive impact on the entire game process, and these nodes will be given higher weights. The transfer attention layer further calculates the importance weights of the transitions between states based on the timing evaluation scores of adjacent nodes and the corresponding state importance weights. This process helps the system accurately evaluate the transition process from the current node to the next node, thereby providing a more accurate reference for path sampling.

[0053] On this basis, the system uses temperature parameters to adjust the path sampling probability to improve the flexibility of the search algorithm. The purpose of adjusting the temperature parameters is to control the exploration scope of the search. When the timing evaluation score of the path sampling does not change significantly, the temperature parameter will be increased, allowing the system to explore more possible paths and avoid local optimal solutions; when the timing evaluation score of the sampled path exceeds the historical optimal, the temperature parameter will be reduced to focus on searching the local area of ​​the optimal path. Through the dynamic adjustment of the temperature parameters, the system can balance global exploration and local search and optimize the selection of strategic paths.

[0054] In order to maintain path diversity, the system evaluates the diversity index of the path by calculating the edit distance between each path in the candidate path set. The edit distance measures the degree of difference between two paths. The greater the difference between the paths, the more different the strategy characteristics are and the higher the path diversity. The system dynamically adjusts the temperature parameter according to the changes in the diversity index. When the diversity index is lower than the target threshold, the temperature parameter increases to expand the search range; conversely, when the diversity index is higher than the target threshold, the temperature parameter decreases to enhance the accuracy of local search. Through this mechanism, the system can ensure the diversity of strategies while avoiding the situation where path selection is overly concentrated on certain similar strategies.

[0055] Each time a new candidate path is generated, the system calculates the weighted sum of the temporal evaluation scores of all tree nodes on the path and uses it as the total evaluation score of the path. If the total evaluation score of the newly generated path exceeds the lowest score in the current candidate path set, the new path is added to the set and the path with the lowest score is removed. This process ensures that the candidate path set always contains the optimal strategy path, thereby providing high-quality path selection for the game model.

[0056] Finally, the system calculates the path weight according to the total evaluation score and sampling probability of each candidate path, and updates the multi-layer game scenario evolution tree according to the path weight. This update process can be continued to dynamically optimize the game behavior model and generate multiple candidate evolution paths with different strategic characteristics, from which the most suitable strategic path for the current game environment is selected.

[0057] In this embodiment, by constructing a time series evaluation function and combining a hierarchical attention mechanism, the multi-layer game scenario evolution tree is dynamically updated, and the game behavior model is effectively optimized. The system can accurately evaluate the importance of each game node according to the change law of the participants' immediate benefits and long-term risks, dynamically adjust the sampling probability of the strategy path, and avoid premature convergence to the local optimal solution. The adaptive adjustment of the temperature parameter enables the system to balance exploration and development, expand the search range and improve the path diversity, thereby improving the flexibility and accuracy of game decisions. By maintaining a set of candidate paths and optimizing them according to the path diversity index, the system ensures that the final selected strategy path has high diversity and adaptability, and can cope with complex and changing game environments. On the whole, the system improves the decision-making efficiency and strategy optimization capabilities of the game model, and can provide effective support and optimization solutions for multi-party games in complex systems.

[0058] In an optional implementation, the candidate evolution path is input into a strategy evaluation model constructed by a spatiotemporal graph neural network to generate a path evaluation index including a technology maturity score, a resource input score, and a strategy benefit score. Selecting the optimal strategy path according to the path evaluation index includes: Constructing a spatiotemporal heterogeneous graph structure, taking technology nodes and resource nodes as heterogeneous graph nodes, setting the technology development sequence relationship as a temporal edge, setting the resource dependency relationship as a spatial edge, constructing a technology node attribute vector based on technology attribute characteristics and timestamp information, constructing a resource node attribute vector based on resource type and capacity information, and combining the technology node attribute vector and resource node attribute vector to form a node feature matrix; The node feature matrix and the temporal edge are input into the causal convolutional network to extract the temporal context features, and the feature weights are calculated through the temporal attention mechanism to obtain enhanced temporal features; the node feature matrix and the spatial edge are input into the graph attention network to extract the node interaction features, and the feature importance is calculated through the spatial attention mechanism to obtain enhanced spatial features; the temporal features and spatial features are input into the gated update unit, the gated weights are calculated, and adaptive fusion is performed to obtain a unified feature representation; Based on the unified feature representation, a multi-layer perceptron and an expert knowledge base are used to perform feature matching to obtain a technical feasibility score; a graph matching network is used to calculate the matching degree between the path resource requirements and the template resources to obtain a resource consumption score; and a graph attention network is used to extract strategy-related features to obtain a strategy value score; Based on the unified feature representation, the mixed Gaussian distribution parameters are predicted, and in combination with the historical evaluation data, the posterior distribution is calculated by variational inference to perform Bayesian calibration on the technical feasibility score, resource consumption score, and strategic value score to obtain a calibrated score; The calibrated score is input into the path optimization objective function, and iterative optimization is performed through a beam search algorithm under the constraints of technical feasibility, resource consumption and strategic value, and finally the optimal strategic path is obtained.

[0059] Exemplarily, first, the system constructs a spatiotemporal heterogeneous graph structure, which contains technology nodes and resource nodes. Technology nodes represent the technical projects or R&D tasks involved in the game scenario, and resource nodes represent the various resources required to realize these technologies, such as funds, equipment, and manpower. The dependencies between technology nodes are represented by temporal edges, which reflect the order of the technology development process; while the dependencies between resource nodes are represented by spatial edges, which indicate the interdependence of different resources. For example, a technology node may depend on a specific resource node, and the availability of resources affects the realization of the technology. Based on these node types, the system constructs a technology node attribute vector for the technology node, which includes technical characteristics such as technology maturity, development complexity, and R&D cycle; at the same time, a resource node attribute vector is constructed for the resource node, which contains information such as resource type, capacity, and service life.

[0060] Next, the system combines the attribute vectors of the technology nodes and resource nodes into a node feature matrix, and inputs the matrix into the causal convolutional network together with the timing edges. The role of the causal convolutional network is to extract temporal context features from time series data, capture the impact of each technology node and resource node over time, and ensure that the system can identify the role of important nodes at different time points. By introducing the temporal attention mechanism, the system can assign different weights to features at different time points according to the changes in time nodes, thereby enhancing the feature impact of important moments and further improving the effectiveness of temporal features.

[0061] In addition, the system inputs the node feature matrix and spatial edges into the graph attention network, which can effectively identify the interactions between technology nodes and resource nodes. Through the spatial attention mechanism, the system assigns different attention weights to nodes according to their relative importance and extracts the most strategic node interaction features. This enables the system to focus on the relationship between key resources and technologies in the spatial dimension, thereby making accurate predictions about interactions in the game environment.

[0062] Then, the obtained temporal features and spatial features are input into the gated update unit, which adaptively fuses these features by calculating the gated weights, and finally generates a unified feature representation. This unified feature representation is a comprehensive reflection of temporal features and spatial features, which not only takes into account the properties of technology nodes and resource nodes, but also incorporates the dynamic interaction characteristics between them, and can fully reflect the multi-dimensional information in the game environment.

[0063] Based on this unified feature representation, the system performs feature matching through multi-layer perceptrons and expert knowledge bases to obtain a technical feasibility score. The technical feasibility score measures the feasibility of a certain strategy path in technical implementation, taking into account factors such as technology maturity, development difficulty, and implementation time. At the same time, the system uses a graph matching network to calculate the match between the path resource requirements and the template resources to obtain a resource consumption score. The resource consumption score reflects the amount of resources required to implement a certain strategy path, helping decision makers understand the resource pressure that the path may bring.

[0064] Through the graph attention network, the system further extracts strategy-related features, generates a strategy value score, and evaluates the contribution of the strategy path to achieving long-term strategy goals. This score takes into account the long-term impact of the strategy, such as market share increase, technological innovation and other factors.

[0065] Subsequently, the system performs Bayesian calibration on the scoring results. First, the system calculates the posterior distribution of the score through variational inference, and uses historical evaluation data to correct the technical feasibility score, resource consumption score, and strategic value score, making these scores more in line with the actual situation, thereby improving the accuracy of the prediction. Through Bayesian calibration, the system can eliminate the bias caused by uncertainty and obtain more reliable evaluation results.

[0066] Finally, the system inputs the calibrated score into the path optimization objective function, and performs iterative optimization through the beam search algorithm under the conditions of meeting the constraints of technical feasibility, resource consumption and strategic value. The beam search algorithm optimizes the strategy selection process through multiple path explorations and finally obtains the optimal strategy path. This optimal path provides decision makers with an optimal implementation plan while considering the balance between technical implementation, resource consumption and strategic value.

[0067] In this embodiment, by constructing a spatiotemporal graph neural network combined with Bayesian inference, the evaluation and optimization capabilities of strategy paths in multi-party games are effectively improved. Through the fusion of temporal features and spatial features, the system can fully capture the complex relationship between technology, resources and strategies in the game process, and then optimize the strategy selection. The Bayesian calibration mechanism further improves the accuracy of the score, making the strategy evaluation more reliable. The beam search algorithm can find the optimal solution in multiple paths, ensuring that the strategy is optimally optimized while meeting the balance of technical feasibility, resource consumption and strategy value. Ultimately, the system can provide decision makers with more accurate and efficient strategy path selection, enhance the system's adaptability and decision-making ability in complex and changing environments, and improve the long-term effect and implementation feasibility of the game strategy.

[0068] In an optional implementation, based on the optimal strategy path, a strategy deployment plan implemented in stages is generated in combination with a knowledge reasoning engine, and the feasibility of the strategy deployment plan is evaluated and optimized through a distributed verification network, including: Constructing a policy deployment knowledge base, the policy deployment knowledge base including a rule base of technology dependency rules, resource allocation rules and progress control rules, and a fact base of acquired technology lists, current available resource status and historical deployment case data; A knowledge reasoning engine is constructed based on the rule base and the fact base. The knowledge reasoning engine extracts the technical dependency relationship of each technical node in the optimal strategy path to obtain a dependency chain. The earliest implementable time of each technical node is calculated by combining the dependency chain and the current system state. The earliest implementable time is matched with the progress control rule in the rule base to generate a time execution window. A resource scheduling plan is calculated based on the time execution window and the resource allocation rule to form a phased deployment plan. Construct a hierarchical verification architecture, which includes a first-level verification component that performs global consistency checks, a second-level verification component that performs technology dependency verification, and a third-level verification component that performs resource constraint verification, and distributes verification tasks to verification components at each level according to a hierarchical authorization mechanism; The verification results of verification components at all levels are input into the Byzantine Fault Tolerant consensus algorithm, and the credibility score of each verification result is calculated based on the reputation weight. The verification results are sorted according to the credibility score, and the highly credible verification opinions are structured to obtain optimization suggestions. Construct a verification indicator vector based on the optimization suggestions, input the verification data of technical feasibility, resource matching and schedule controllability into the corresponding indicator calculation model, and generate a quantitative verification evaluation result; The verification and evaluation results are input into the policy optimization network to generate targeted optimization and adjustment plans, which are applied to the execution plan for updating. The updated execution plan is re-input into the hierarchical verification architecture for evaluation. When the verification and evaluation results meet the preset standards, the final policy deployment plan is output. When the verification and evaluation results do not meet the standards, the execution plan is returned to the policy optimization network for iterative optimization.

[0069] Exemplarily, first, a policy deployment knowledge base is constructed. The knowledge base consists of two parts: a rule base and a fact base. The rule base contains technology dependency rules, such as "technology A can only be deployed after technology B is deployed"; resource allocation rules, such as "technology C requires Y units of X-type resources"; and progress control rules, such as "technology D must be deployed before Z time". The fact base contains a list of technologies that have been mastered, such as technologies A, B, and C; the current available resource status, such as 10 units of X-type resources and 5 units of Y-type resources; and historical deployment case data, such as the average deployment time of technology A in past cases is 3 days.

[0070] Next, a knowledge reasoning engine is built based on the knowledge base. Take the deployment of a strategy that includes technologies A, B, C, and D as an example. First, the reasoning engine extracts the dependency relationship of each technology node based on the rule base, for example, technology C depends on technologies A and B. After obtaining the dependency chain, the earliest implementable time of each technology node is calculated in combination with the current system status (mastered technologies, available resources). For example, if the deployment time of technologies A and B is 1 day and 2 days respectively, then the earliest implementable time of technology C is the 3rd day. Then, the earliest implementable time is matched with the progress control rules to generate a time execution window. For example, if technology C must be deployed before the 5th day, then its time execution window is from the 3rd day to the 5th day. Finally, the resource scheduling plan is calculated based on the time execution window and resource allocation rules, for example, 2 units of X-type resources are allocated to technology C on the 3rd day, and finally a phased deployment plan is formed, such as deploying technologies A and B in the first phase, deploying technology C in the second phase, and deploying technology D in the third phase.

[0071] Then, a hierarchical verification architecture is constructed. The architecture contains three levels of verification components: the first level performs global consistency checks, such as checking whether the entire deployment plan meets the overall goal; the second level performs technical dependency verification, such as checking whether the deployment of technology C meets its dependencies; the third level performs resource constraint verification, such as checking whether resource allocation exceeds the limit of available resources. Suppose the first-level verification component finds that the deployment plan deviates from the overall goal, the second-level verification component finds that the dependencies of technology C are not fully met, and the third-level verification component finds that the resource allocation plan is feasible. These verification tasks will be issued to the corresponding verification components according to the hierarchical authorization mechanism.

[0072] The verification results of the verification components at all levels will be input into the Byzantine Fault Tolerant consensus algorithm. The algorithm calculates the credibility score of each verification result based on the credibility weight. For example, the credibility of the first-level verification component is 0.8, the second-level is 0.9, and the third-level is 0.7. Assume that the first-level verification result is "inconsistent", the second-level is "partial dependency satisfaction", and the third-level is "resource satisfaction". According to the credibility weighted calculation, the final credibility scores are 0.8*(-1)=-0.8, 0.9*0.5=0.45, and 0.7*1=0.7. The verification results are sorted according to the credibility score to obtain the third-level, second-level, and first-level rankings. The high-credibility verification opinions are structured to obtain optimization suggestions, such as "giving priority to satisfying the dependency of technology C".

[0073] According to the optimization suggestions, a verification indicator vector is constructed, and the verification data of technical feasibility, resource matching and schedule controllability are input into the corresponding indicator calculation model to generate quantitative verification evaluation results. For example, the technical feasibility score is 0.8, the resource matching score is 0.9, and the schedule controllability score is 0.7.

[0074] The verification and evaluation results are input into the policy optimization network to generate targeted optimization and adjustment plans. For example, the deployment time of technology C is advanced to the second day. The optimization and adjustment plan is applied to the execution plan for update, and the updated execution plan is re-input into the hierarchical verification architecture for evaluation. If the verification and evaluation results meet the preset standards (for example, the scores of each indicator are greater than 0.8), the final policy deployment plan is output. If it does not meet the standards, the execution plan is returned to the policy optimization network for iterative optimization until the preset standards are met.

[0075] In this embodiment, the deployment plan is automatically generated by the knowledge reasoning engine, and the plan is quickly evaluated and optimized through the distributed verification network, which can significantly shorten the policy deployment cycle. The hierarchical verification architecture can fully verify the deployment plan from multiple dimensions, and the Byzantine fault-tolerant consensus algorithm can effectively identify and filter erroneous verification results, thereby improving the accuracy and reliability of the deployment plan. The iterative optimization mechanism can dynamically adjust the deployment plan according to the actual situation, so as to better adapt to the complex and changing deployment environment.

[0076] Figure 2 FIG. 1 is a schematic diagram of a system for dynamically evolving and optimizing game behaviors in space according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to obtain historical behavior data of participants in the space field, perform time series feature analysis on the historical behavior data through a deep neural network, obtain the periodic change law of the participants' behavior, and build a feature extraction model based on the periodic change law using a distributed learning framework to generate a multi-dimensional feature vector. A dynamic relationship diagram of the participants is built according to the multi-dimensional feature vector, and the dynamic relationship diagram is input into a preset multi-agent reinforcement learning model. The multi-agent reinforcement learning model adopts an encoder structure and a multi-layer self-attention mechanism to output a game behavior model including an adaptive utility function and dynamic decision-making rules; The second unit is used to construct a multi-layer game scenario evolution tree according to the game behavior model, wherein each tree node contains the state information and decision space of the participants, input the state information and decision space into the dual neural network structure, respectively calculate the immediate benefit value and long-term risk value of the participants in each game scenario, and construct an evaluation function based on the temporal variation law of the immediate benefit value and the long-term risk value. The multi-layer game scenario evolution tree is dynamically updated by using a search algorithm with a hierarchical attention mechanism. The hierarchical attention mechanism adaptively adjusts the search strategy according to the evaluation function, determines the sampling weight of the evolution path, and generates multiple candidate evolution paths with different strategy characteristics; The third unit is used to input the candidate evolution path into the strategy evaluation model constructed by the spatiotemporal graph neural network, generate path evaluation indicators including technology maturity score, resource input score and strategy benefit score, select the optimal strategy path according to the path evaluation indicators, and generate a strategy deployment plan for phased implementation based on the optimal strategy path in combination with the knowledge reasoning engine, and evaluate and optimize the feasibility of the strategy deployment plan through a distributed verification network.

[0077] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0078] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0079] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamic evolution and strategy deduction optimization of game behavior in space, characterized in that: include: Obtain historical behavior data of participants in the space field, perform time series feature analysis on the historical behavior data through a deep neural network, obtain the periodic change law of the participants' behavior, build a feature extraction model based on the periodic change law using a distributed learning framework, generate a multi-dimensional feature vector, build a dynamic relationship graph of the participants based on the multi-dimensional feature vector, input the dynamic relationship graph into a preset multi-agent reinforcement learning model, and the multi-agent reinforcement learning model adopts an encoder structure and a multi-layer self-attention mechanism to output a game behavior model containing an adaptive utility function and dynamic decision-making rules; A multi-layer game scenario evolution tree is constructed according to the game behavior model, wherein each tree node contains the state information and decision space of the participants, the state information and decision space are input into the dual neural network structure, and the immediate benefit value and long-term risk value of the participants in each game scenario are respectively calculated. Based on the temporal variation law of the immediate benefit value and the long-term risk value, an evaluation function is constructed, and a search algorithm with a hierarchical attention mechanism is used to dynamically update the multi-layer game scenario evolution tree. The hierarchical attention mechanism adaptively adjusts the search strategy according to the evaluation function, determines the sampling weight of the evolution path, and generates multiple candidate evolution paths with different strategy characteristics; The candidate evolution path is input into the strategy evaluation model constructed by the spatiotemporal graph neural network to generate path evaluation indicators including technology maturity score, resource input score and strategy benefit score, the optimal strategy path is selected according to the path evaluation indicators, and based on the optimal strategy path, a strategy deployment plan for phased implementation is generated in combination with the knowledge reasoning engine, and the feasibility of the strategy deployment plan is evaluated and optimized through a distributed verification network.

2. The method according to claim 1, characterized in that A dynamic relationship diagram of the participants is constructed according to the multi-dimensional feature vector, and the dynamic relationship diagram is input into a preset multi-agent reinforcement learning model. The multi-agent reinforcement learning model adopts an encoder structure and a multi-layer self-attention mechanism, and outputs a game behavior model including an adaptive utility function and dynamic decision rules, including: Generate an initial dynamic relationship graph based on the multidimensional feature vector, where nodes represent participants and graph edges represent the interaction relationship between participants; embed nodes with features through a deep neural network to obtain node representation vectors, and use a temporal convolutional network to extract interaction history to obtain dynamic pattern features; combine the node representation vector with the dynamic pattern features to calculate edge weights, and update the topological structure of the dynamic relationship graph based on the edge weights and node representation vectors; The dynamic relationship graph is input into a multi-agent reinforcement learning model, wherein the multi-agent reinforcement learning model comprises an encoder network and a strategy network; the encoder network encodes the dynamic relationship graph through a node feature layer, a relationship feature layer and a time series feature layer to obtain a state representation vector; the strategy network generates an action probability distribution of each agent based on the state representation vector; Constructing a multi-agent value network, predicting the state value and action value of each agent based on the state representation vector and the action probability distribution; inputting the state value and action value into the utility function generation module to construct a utility function including individual utility and group utility; Calculate the reward signal of each agent based on the utility function, and input the reward signal into the policy optimization module; the policy optimization module uses the policy gradient method to update the policy network parameters to achieve dynamic optimization of the agent decision rules; Adopting the experience replay mechanism to store the interaction data of the intelligent agents, and synchronously updating the value network and utility function parameters based on the interaction data; realizing the joint optimization of the strategy network, value network and utility function through multi-agent collaborative learning; The optimized strategy network, value network and utility function are integrated into a model to output a game behavior model that includes an adaptive utility function and dynamic decision rules.

3. The method according to claim 2, characterized in that Constructing a multi-agent value network, predicting the state value and action value of each agent based on the state representation vector and action probability distribution; inputting the state value and action value into the utility function generation module, and constructing a utility function containing individual utility and group utility includes: Construct a state value network and an action value network. The state value network adopts a three-layer fully connected network structure, and the state representation vector is input to obtain the state value prediction value; the action value network adopts a double-branch structure, the first branch processes the state representation vector, and the second branch processes the action probability distribution. The two branch features are fused through the outer product operation and then the action value prediction value is obtained through the fully connected layer; Calculate a time series difference target value, the time series difference target value is calculated based on the immediate reward value, the state value prediction value and the action value prediction value; construct a state value loss function by the difference between the time series difference target value and the state value prediction value, and construct an action value loss function by the difference between the time series difference target value and the action value prediction value; update the value network parameters based on the state value loss function and the action value loss function to obtain optimized state value and action value; The optimized action value is used to represent the individual utility term, and the group utility term is obtained by weighted calculation of the optimized state value and the action value of the neighborhood agent; an initial utility function is constructed by combining the individual utility term and the group utility term; The gradient value of the initial utility function with respect to the weight coefficient is calculated, and based on the gradient value, the weight coefficients of the individual utility term and the group utility term are updated using the gradient descent method to obtain the final utility function.

4. The method according to claim 1, characterized in that: A multi-layer game scenario evolution tree is constructed according to the game behavior model, wherein each tree node contains the state information and decision space of the participants, and the state information and decision space are input into the dual neural network structure to calculate the immediate benefit value and long-term risk value of the participants in each game scenario respectively, including: A multi-layer game scenario evolution tree is constructed based on the dynamic decision rules in the game behavior model, and the state vectors and decision space information of the participants are stored in the tree nodes; the transition probability matrix between adjacent nodes is calculated based on the state transition function; Monte Carlo sampling is performed according to the transition probability matrix to generate an evolution path from the root node to the leaf node; the node sequence corresponding to the evolution path and the historical state sequence of the participants are stored as feature inputs in each tree node; Construct a dual neural network structure, including a graph attention network and a hierarchical recurrent neural network, input the state vector in the tree node into the graph attention network, generate a feature representation matrix of the participants through the feature transformation layer, calculate the attention coefficient between the participants based on the feature representation matrix, multiply the attention coefficient with the feature representation matrix to obtain a weighted feature, and perform a residual connection between the weighted feature and the feature representation matrix, transform the residually connected features through a multi-head attention mechanism, extract the correlation features between the participants, and input the correlation features into the adaptive pooling layer of the graph attention network for dimensionality reduction aggregation to obtain an instantaneous benefit value reflecting the game state of the participants at the current node; The historical state sequences stored in the tree nodes are grouped according to the participant identifiers, and the grouped sequences are input into a hierarchical recurrent neural network. The recurrent neural network selectively memorizes the grouped sequences through a gating unit, outputs the state features of each time step, calculates the attention weight of the time dimension for the state features, and obtains a temporal correlation representation that describes the state evolution law. At the same time, the attention weight of the participant dimension is calculated for the state features to obtain an interactive correlation representation that describes the game relationship. The temporal correlation representation and the interactive correlation representation are weightedly combined with the state features, and the long-term risk value reflecting the cumulative effect of the participants on the evolution path is output through the prediction module of the hierarchical recurrent neural network.

5. The method according to claim 1, characterized in that Based on the temporal variation law of the instant return value and the long-term risk value, an evaluation function is constructed, and a search algorithm with a hierarchical attention mechanism is used to dynamically update the multi-layer game scenario evolution tree. The hierarchical attention mechanism adaptively adjusts the search strategy according to the evaluation function, determines the sampling weight of the evolution path, and generates multiple candidate evolution paths with different strategy characteristics, including: Calculate the rate of change of benefits according to the instant benefit value of the game state of adjacent nodes, and calculate the risk change trend according to the long-term risk value of the historical state sequence; construct a time series evaluation function by weighted combination of the instant benefit value, the rate of change of benefits and the risk change trend; calculate the time series evaluation score of each tree node in the multi-layer game scenario evolution tree according to the time series evaluation function; A state attention layer is constructed based on the temporal evaluation scores of tree nodes. The state importance weights of tree nodes are output through the state attention layer. At the same time, the temporal evaluation scores and corresponding state importance weights of adjacent tree nodes are input into the transfer attention layer to obtain the importance weights of state transfers. The complete path sampling probability from the root node to the leaf node is calculated based on the state importance weights of the tree nodes and the importance weights of state transfers. The temperature parameter is introduced to adjust the path sampling probability, and the gradient change value of the timing evaluation function is detected. When the gradient change value is less than the convergence threshold, the temperature parameter is increased, and the timing evaluation score of the sampling path is detected. When the score exceeds the historical optimal, the temperature parameter is reduced. The ratio of the path sampling probability to the temperature parameter is used as the adjusted sampling probability, and the candidate evolution path is generated based on the adjusted sampling probability. Maintain a fixed-size candidate path set, calculate the weighted sum of the temporal evaluation scores of all tree nodes on the newly generated path as the total evaluation score of the path, and when the total evaluation score of the new path is higher than the lowest score in the set, add the new path to the candidate path set and remove the path with the lowest score; The edit distance between paths in the candidate path set is calculated to obtain the path diversity index, and the temperature parameter is dynamically adjusted based on the diversity index. When the diversity index is lower than the target threshold, the temperature parameter is increased to expand the search range. When the diversity index is higher than the target threshold, the temperature parameter is reduced to strengthen the local search. The path weight is calculated according to the total evaluation score and sampling probability of each candidate path, and the multi-layer game scenario evolution tree is updated based on the path weight, and finally multiple candidate evolution paths with different strategic characteristics are obtained.

6. The method according to claim 1, characterized in that The candidate evolution path is input into the strategy evaluation model constructed by the spatiotemporal graph neural network to generate a path evaluation index including a technology maturity score, a resource input score and a strategy benefit score. The optimal strategy path is selected according to the path evaluation index, including: Constructing a spatiotemporal heterogeneous graph structure, taking technology nodes and resource nodes as heterogeneous graph nodes, setting the technology development sequence relationship as a temporal edge, setting the resource dependency relationship as a spatial edge, constructing a technology node attribute vector based on technology attribute characteristics and timestamp information, constructing a resource node attribute vector based on resource type and capacity information, and combining the technology node attribute vector and resource node attribute vector to form a node feature matrix; The node feature matrix and the temporal edge are input into the causal convolutional network to extract the temporal context features, and the feature weights are calculated through the temporal attention mechanism to obtain enhanced temporal features; the node feature matrix and the spatial edge are input into the graph attention network to extract the node interaction features, and the feature importance is calculated through the spatial attention mechanism to obtain enhanced spatial features; the temporal features and spatial features are input into the gated update unit, the gated weights are calculated, and adaptive fusion is performed to obtain a unified feature representation; Based on the unified feature representation, a multi-layer perceptron and an expert knowledge base are used to perform feature matching to obtain a technical feasibility score; a graph matching network is used to calculate the matching degree between the path resource requirements and the template resources to obtain a resource consumption score; and a graph attention network is used to extract strategy-related features to obtain a strategy value score; Based on the unified feature representation, the mixed Gaussian distribution parameters are predicted, and in combination with the historical evaluation data, the posterior distribution is calculated by variational inference to perform Bayesian calibration on the technical feasibility score, resource consumption score, and strategic value score to obtain a calibrated score; The calibrated score is input into the path optimization objective function, and iterative optimization is performed through a beam search algorithm under the constraints of technical feasibility, resource consumption and strategic value, and finally the optimal strategic path is obtained.

7. The method according to claim 1, characterized in that Based on the optimal strategy path, a strategy deployment plan implemented in stages is generated in combination with a knowledge reasoning engine, and the feasibility of the strategy deployment plan is evaluated and optimized through a distributed verification network, including: Constructing a policy deployment knowledge base, the policy deployment knowledge base including a rule base of technology dependency rules, resource allocation rules and progress control rules, and a fact base of acquired technology lists, current available resource status and historical deployment case data; A knowledge reasoning engine is constructed based on the rule base and the fact base. The knowledge reasoning engine extracts the technical dependency relationship of each technical node in the optimal strategy path to obtain a dependency chain. The earliest implementable time of each technical node is calculated by combining the dependency chain and the current system state. The earliest implementable time is matched with the progress control rule in the rule base to generate a time execution window. A resource scheduling plan is calculated based on the time execution window and the resource allocation rule to form a phased deployment plan. Construct a hierarchical verification architecture, which includes a first-level verification component that performs global consistency checks, a second-level verification component that performs technology dependency verification, and a third-level verification component that performs resource constraint verification, and distributes verification tasks to verification components at each level according to a hierarchical authorization mechanism; The verification results of verification components at all levels are input into the Byzantine Fault Tolerant consensus algorithm, and the credibility score of each verification result is calculated based on the reputation weight. The verification results are sorted according to the credibility score, and the highly credible verification opinions are structured to obtain optimization suggestions. Construct a verification indicator vector based on the optimization suggestions, input the verification data of technical feasibility, resource matching and schedule controllability into the corresponding indicator calculation model, and generate a quantitative verification evaluation result; The verification and evaluation results are input into the policy optimization network to generate targeted optimization and adjustment plans, which are applied to the execution plan for updating. The updated execution plan is re-input into the hierarchical verification architecture for evaluation. When the verification and evaluation results meet the preset standards, the final policy deployment plan is output. When the verification and evaluation results do not meet the standards, the execution plan is returned to the policy optimization network for iterative optimization.

8. A system for dynamic evolution and strategy deduction optimization of game behavior in space, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain historical behavior data of participants in the space field, perform time series feature analysis on the historical behavior data through a deep neural network, obtain the periodic change law of the participants' behavior, and build a feature extraction model based on the periodic change law using a distributed learning framework to generate a multi-dimensional feature vector. A dynamic relationship diagram of the participants is built according to the multi-dimensional feature vector, and the dynamic relationship diagram is input into a preset multi-agent reinforcement learning model. The multi-agent reinforcement learning model adopts an encoder structure and a multi-layer self-attention mechanism to output a game behavior model including an adaptive utility function and dynamic decision-making rules; The second unit is used to construct a multi-layer game scenario evolution tree according to the game behavior model, wherein each tree node contains the state information and decision space of the participants, input the state information and decision space into the dual neural network structure, respectively calculate the immediate benefit value and long-term risk value of the participants in each game scenario, and construct an evaluation function based on the temporal variation law of the immediate benefit value and the long-term risk value. The multi-layer game scenario evolution tree is dynamically updated by using a search algorithm with a hierarchical attention mechanism. The hierarchical attention mechanism adaptively adjusts the search strategy according to the evaluation function, determines the sampling weight of the evolution path, and generates multiple candidate evolution paths with different strategy characteristics; The third unit is used to input the candidate evolution path into the strategy evaluation model constructed by the spatiotemporal graph neural network, generate path evaluation indicators including technology maturity score, resource input score and strategy benefit score, select the optimal strategy path according to the path evaluation indicators, and generate a strategy deployment plan for phased implementation based on the optimal strategy path in combination with the knowledge reasoning engine, and evaluate and optimize the feasibility of the strategy deployment plan through a distributed verification network.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images