Multi-dimensional resource dynamic game optimization control method and system based on multi-agent system

Abstract

The present invention discloses a multi-dimensional resource dynamic game optimization control method and system based on a multi-agent system, which specifically relates to the technical field of multi-agent system decision-making control, and includes the following steps: constructing a tensor-form resource coupling map with resource types as nodes and dynamic weights as edges; collecting all agent strategy trajectories, calculating the group strategy divergence index and the resource coupling tension index, inputting the trained game prediction model to obtain a convergence risk score, and adjusting the individual or group strategy if the score exceeds the threshold; updating the map and behavior data after adjustment to form an adaptive feedback closed loop, thereby improving the strategy convergence and resource allocation coordination; the present invention realizes resource relationship modeling, two-dimensional diagnosis of strategy behavior and resource conflict, and an adaptive optimization closed loop of behavior and resource linkage by constructing a resource coupling map, calculating dual-index indicators and introducing a risk scoring mechanism, thereby improving the stability and coordination efficiency of the multi-agent system.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-agent system decision control, and more specifically, to a multi-dimensional resource dynamic game optimization control method and system based on a multi-agent system. Background Art

[0002] In recent years, multi-agent systems, driven by intelligent agents and operating autonomously, have been widely used in scenarios such as intelligent manufacturing, the Internet of Vehicles, data center resource scheduling, and robotic swarm control due to their distributed, scalable, and environmentally adaptable nature. In such systems, each intelligent agent typically possesses the ability to autonomously evolve its strategy and make parallel decisions across multiple heterogeneous resources. However, as the system scale expands and the resource dimension becomes more complex, the strategic behavior and resource competition among multiple agents often exhibit highly nonlinear and dynamically coupled game relationships, which can easily lead to policy divergence, frequent resource conflicts, and reduced overall system efficiency.

[0003] Traditional resource scheduling methods or policy optimization mechanisms often only consider static resource allocation rules or optimal response strategies based on a single agent. They are unable to effectively capture the evolutionary trends of group behavior in multi-agent systems, and are also unable to model the dynamic coordination and conflict relationships between resources that change over time. They lack overall behavior predictability and system stability guarantees.

[0004] Especially in environments with multi-dimensional heterogeneous resources, such as computing resources, communication resources, energy resources, and storage resources, different resource types exhibit both synergistic relationships (for example, computing and communication work together to achieve efficient task completion) and inhibitory relationships (for example, energy consumption constraints inhibit high-frequency computing operations). Traditional models are unable to dynamically identify the evolutionary paths of these relationships. Furthermore, the interactions between multi-agent strategies often manifest as implicit deviations and confrontations. Traditional methods often cannot accurately identify whether group behavior is in an unbalanced state and lack efficient risk identification mechanisms. Therefore, this paper proposes a multi-dimensional resource dynamic game optimization control method and system based on a multi-agent system to address these issues. Summary of the Invention

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The multi-dimensional resource dynamic game optimization control method based on a multi-agent system includes the following steps:

[0007] Step 1: Construct a multi-dimensional heterogeneous resource coupling graph. By modeling the synergy and inhibition relationships between resource types within the same control cycle, a graph structure is formed with resource types as nodes and dynamic weights as edges, forming a tensor-like graph for resource state evolution.

[0008] Step 2: Collect the strategy trajectories of all intelligent agents over multiple consecutive control cycles, calculate the group strategy divergence index and the group resource coupling tension index based on the dynamic changes of edge weights in the resource coupling graph, and use them as dual index indicators to characterize the overall behavior deviation and resource conflict status of the multi-agent system;

[0009] Step 3: Input the dual index indicator into the trained game optimization prediction model. The model outputs a strategy behavior convergence risk score. If the score is below the first threshold, the strategy remains unchanged. If the score is above the first threshold, the detailed analysis process of step 4 is entered.

[0010] Step 4: Evaluate the strategy trajectory and resource behavior of all intelligent agents individually, calculate the individual strategy deviation index of each agent, and compare this index with the second threshold. If the proportion of agents exceeding the second threshold among all agents is lower than the set ratio, adjust the strategy evolution direction and resource selection behavior only for the agents exceeding the second threshold; if the proportion is not lower than the set ratio, perform a unified strategy update operation for all agents;

[0011] Step 5: After the strategy adjustment or unified update is completed, the edge weight parameters and agent behavior trajectory data in the resource coupling graph are updated. This serves as the basis for the next round of evaluation and optimization, forming a continuously iterative dynamic feedback loop to achieve adaptive convergence of strategy evolution and global optimization of resource allocation behavior.

[0012] In a preferred embodiment, in the process of constructing a multi-dimensional heterogeneous resource coupling map, resource types are limited to at least computing resources, storage resources, communication resources, and energy consumption resources. In each control cycle, the usage intensity data of each type of resource is collected as raw observation values, and the raw observation values ​​are used to calculate the coordination function and inhibition function between resource types.

[0013] The synergy function is used to reflect the degree of positive mutual influence between two resource types in actual tasks, and the inhibition function is used to reflect the negative constraint relationship between two resource types in allocation behavior;

[0014] The calculation results of the cooperation function and the inhibition function are combined into a composite weight value, and the composite weight value is used as the dynamic weight of the edge to connect the corresponding resource type nodes, thereby constructing a graph structure with resource types as nodes and dynamic weights as edges. A timing mark is added to the graph structure to record the evolution of the structure in multiple consecutive control cycles.

[0015] In a preferred embodiment, in the process of constructing a multi-dimensional heterogeneous resource coupling map, the calculation method of the cooperation function and the inhibition function between resource types is based on the resource synchronization utilization rate and conflict occupancy rate within the control cycle, specifically:

[0016] In each control cycle, the number of resource allocation events in which each resource type participates in each intelligent agent behavior is counted, and the joint usage frequency of different resource types being simultaneously called in the same time period is calculated. The initial value of the coordination function is obtained by dividing this joint usage frequency by the geometric mean of the individual usage frequencies of each resource.

[0017] By comparing the number of allocation conflict events between different resource types, that is, the number of times two resource types cannot be allocated simultaneously due to quota restrictions in resource allocation decisions, and combining the weights of task failure events caused by agent behavior, the initial value of the inhibition function of one resource on another is calculated;

[0018] The synergy function and the inhibition function are combined into the dynamic weight of the edge through normalized mapping, and the edge weight between resource types in the graph structure is updated to form the edge weight parameters of the tensor form graph that reflects the evolution of resource status.

[0019] In a preferred embodiment, in the process of calculating the group strategy divergence index, the strategy trajectories of all intelligent agents in multiple consecutive control cycles are first time-series encoded, and each strategy trajectory is represented as a set of vectors of fixed length. In the same control cycle, the mean Euclidean distance between the strategy vector of each agent and the strategy vectors of all other agents is calculated, and the divergence score of the agent is obtained by the ratio of the mean to the standard deviation of the change of the agent strategy vector itself; then the divergence scores of all agents are normalized and mean aggregated to obtain the group strategy divergence index, which is used to measure the offset amplitude, behavioral divergence degree and consistency of evolutionary trend of all intelligent agents in strategy behavior.

[0020] In a preferred embodiment, in the process of calculating the group resource coupling tension index based on the dynamic change of edge weights in the resource coupling graph, the weight change values ​​of all edges are first extracted within two adjacent control cycles, and the tension strength of each edge during the change process is calculated. The calculation method is to multiply the square of the weight difference between the current cycle and the previous cycle of the edge by the initial weight factor; then, with each resource type as the center, the tension strength values ​​of all its connected edges are summarized to calculate the local tension response value of the resource type; then, the weighted average of the local tension response values ​​of all resource types is taken to generate the group resource coupling tension index, which is used to measure the degree of structural tension occurring during the evolution of resource states within the current control cycle and the resource coordination stability at the system level.

[0021] In a preferred embodiment, when inputting the dual-index indicator into a trained game optimization prediction model and having the model output a strategy behavior convergence risk score, the group strategy divergence index and the group resource coupling tension index are first normalized and mapped, and a dual-index input vector is formed according to a preset dimension. This input vector is input into the game optimization prediction model as an expression of the current multi-agent system behavior state;

[0022] This game optimization prediction model is trained based on historical evolution data within the control cycle, and adopts a nonlinear prediction structure constructed by integrating a feedforward neural network and a graph association reasoning structure. By fitting and predicting the positional relationship of the input vector in the historical behavior mapping space, it outputs a convergence risk score for the current behavior state in the strategy evolution process. This score is used to reflect the behavioral consistency, evolutionary stability and future convergence trend of the current multi-agent system in the strategy dimension, and is used to determine whether to trigger the next step of the detailed analysis process.

[0023] In a preferred embodiment, in the process of evaluating the policy trajectories and resource behaviors of all intelligent agents separately and calculating the individual policy deviation index of each agent, the policy trajectory of each agent in multiple consecutive control cycles is first selected, and a sequence of its policy behavior vectors in each control cycle is constructed;

[0024] After normalizing the vector sequence, the change amplitude of the strategy behavior vector between adjacent control cycles is obtained, and the average change amplitude of the agent is used as the internal variability of its strategy trajectory;

[0025] Select the strategy behavior vector sequence of other agents in all intelligent agents, and construct the strategy average vector difference of this agent relative to all other agents as the external difference of its behavior deviation degree;

[0026] The internal difference between the internal volatility of the strategy trajectory and the external difference between the behavioral deviation is synthesized into a single indicator in the form of a weighted linear combination, and the indicator is smoothed within a fixed window to finally generate the individual strategy deviation index of the agent at the current stage. It is used to measure the instability of the agent's strategic behavior, the difference from the group behavior, and the potential abnormal behavioral trend.

[0027] In a preferred embodiment, when adjusting the strategy evolution direction and resource selection behavior only for the exceeding agent, i.e., the agent exceeding the second threshold, a behavior direction correction factor is set based on the difference direction between the strategy behavior vector of the exceeding agent in the current control cycle and the strategy behavior mean vector of all intelligent agents. The correction factor is the negative of the cosine value of the angle between the difference vector and the unit convergence vector multiplied by a non-zero proportional coefficient γ. The correction factor is used in the strategy generation vector superposition term in the next control cycle of the exceeding agent to adjust the direction of the strategy behavior evolution path;

[0028] The resource set used by the exceeding agent in the current control cycle is collected, and the resources with the largest difference from the resource usage frequency ranking of all current agents constitute the deviation resource set. The resource allocation priority compression factor η of the exceeding agent in the next cycle is set to the initial resource selection probability multiplied by [1-fd / (ft+C)], where fd is the resource frequency used by the agent, ft is the group average resource usage frequency, and C is a preset non-zero constant. The compression operation is set to only act on the resource types in the deviation resource set, so as to achieve dynamic constraints on resource selection behavior.

[0029] In a preferred embodiment, if the number of agents exceeding the second threshold accounts for no less than a set proportion of all agents, a unified policy update operation for all agents is performed, including extracting the mean vector of the policy trajectories of all agents in the past N control cycles as a baseline behavior model, obtaining the average offset of all agent behavior vectors in the model, and taking the policy behavior vector with the smallest offset as the convergence direction vector, resetting the initial values ​​of the policy generation parameters of all agents to the convergence direction vector multiplied by the initialization scale factor, and sorting all resource behavior parameters according to the resource selection frequency, mapping them to the preferred resource interval constructed based on the probability of group resource co-occurrence, and uniformly applying the mapping structure in the new control cycle to achieve synchronous convergence of agent behavior and resource selection behavior, thereby improving overall behavior consistency and resource scheduling coordination.

[0030] In a preferred embodiment, a multi-dimensional resource dynamic game optimization control system based on a multi-agent system includes:

[0031] The resource graph builder is used to construct a multi-dimensional heterogeneous resource coupling graph. By modeling the collaborative and inhibitory relationships between resource types within the same control cycle, it generates a graph structure with resource types as nodes and dynamic weights as edges, and outputs a tensor-based graph to describe the evolution of resource states.

[0032] A behavioral index calculator collects the strategy trajectories of all intelligent agents over multiple consecutive control cycles, calculates the group strategy divergence index and the group resource coupling tension index obtained by the dynamic changes of edge weights in the resource coupling graph, and generates a dual-index indicator to characterize the overall behavioral deviation degree and resource conflict status of the multi-agent system;

[0033] A game risk discriminator is used to receive the dual index indicator and input it into the trained game optimization prediction model, output the strategy behavior convergence risk score, and determine whether to trigger the subsequent analysis and adjustment process based on the comparison result of the score with the first threshold;

[0034] The local evolution controller is used to evaluate the strategy trajectory and resource behavior of all intelligent agents separately, calculate the individual strategy deviation index, and based on the comparison result of the index with the second threshold, determine whether the proportion of agents exceeding the standard is lower than the set ratio. If it is lower, only the strategy evolution direction and resource selection behavior of the exceeding agents are adjusted. If it is not lower, a unified strategy update operation is performed on all agents.

[0035] The evolutionary feedback updater is used to update the edge weight parameters and agent behavior trajectory data in the resource coupling graph after the strategy adjustment or unified update is completed, providing a new input basis for the next control cycle and forming an adaptive convergence feedback loop for strategy evolution.

[0036] Technical effects and advantages of the present invention:

[0037] This paper models the collaborative and inhibitory relationships between multidimensional heterogeneous resources by constructing a graph structure with resource types as nodes and dynamic weights as edges. This graph structure is then converted into a tensor-like graph as the fundamental representation of resource state evolution. This approach allows for the explicit expression and dynamic tracking of complex dependencies between resources, improving the parseability of resource modeling and enhancing the system's ability to perceive structural changes at the resource level, thereby facilitating more accurate assessment and adjustment of subsequent policy optimization processes at the resource level.

[0038] This method collects the strategy trajectories of all intelligent agents over multiple consecutive control cycles, calculates the group strategy divergence index and the group resource coupling tension index based on the dynamic changes in the edge weights of the resource coupling graph, and uses these two as dual-index indicators to jointly input into a trained game optimization prediction model. This achieves simultaneous assessment and prediction of the overall behavioral deviation degree and resource conflict status of the multi-agent system. This dual-index mechanism breaks through the limitations of single-behavior judgment or resource monitoring, enabling the system to dynamically perceive whether the current game situation is at risk of non-convergence, providing a scientific basis and quantifiable trigger conditions for subsequent detailed analysis and behavioral intervention.

[0039] During the strategy evolution process, the present invention designs a mechanism that triggers a detailed analysis process based on the strategy behavior convergence risk score. Based on whether the individual strategy deviation index exceeds the threshold and the proportion of agents exceeding the threshold, local behavior adjustments or overall strategy updates are performed. After completing the behavior adjustment, the system synchronously updates the edge weight parameters and agent behavior trajectory data in the resource coupling graph, thus forming a complete behavior-resource-feedback closed-loop optimization path. This mechanism significantly improves the adaptability and convergence of strategy evolution, and can continuously maintain the synergy and resource scheduling efficiency of the multi-agent system in a dynamic environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0041] Figure 1 This is a schematic diagram of the multi-dimensional resource dynamic game optimization control method based on a multi-agent system in the present invention.

[0042] Figure 2 This is a schematic diagram of the multi-dimensional resource dynamic game optimization control system based on a multi-agent system in the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] Reference Figure 1 - Figure 2 The following examples were obtained:

[0045] Example 1: A multi-dimensional resource dynamic game optimization control method based on a multi-agent system, comprising the following steps:

[0046] Step 1: Construct a multidimensional heterogeneous resource coupling graph. By modeling the synergy and inhibition relationships between resource types within the same control cycle, a graph structure is formed with resource types as nodes and dynamic weights as edges, forming a tensor-like graph for resource state evolution. Constructing a multidimensional heterogeneous resource coupling graph involves establishing a structured representation model that describes the interactions between resources based on their usage within the same control cycle during the operation of the multi-agent system. In this step, each resource type is abstracted as a node in the graph structure, and edges between them are constructed based on the degree of synergy and inhibition between resource types. The weight of each edge is dynamically updated based on observed resource usage data, allowing edge weights to evolve over time. This graph structure is adjusted based on real-time data within each control cycle to accurately reflect the coupling strength and directionality between the current resource states. Finally, this graph structure is encoded as a tensor for subsequent processing and model input, serving as the input basis and behavioral context for resource state evolution.

[0047] Step 2: Collect the policy trajectories of all intelligent agents over multiple consecutive control cycles and calculate the group policy divergence index and the group resource coupling tension index, based on the dynamic changes in edge weights in the resource coupling graph. These dual-index indicators are used to characterize the overall behavioral deviation and resource conflict status of the multi-agent system. Collecting the policy trajectories of all intelligent agents over multiple consecutive control cycles involves collecting the policy behavior data of each agent over a continuous time period at the system level and converting it into a set of policy trajectories expressed as a time series. After obtaining all policy trajectories, a statistical analysis is first performed at the group behavior level. By comparing the policy paths of each agent, an overall indicator of the degree of consistency and divergence between agent behaviors is calculated, known as the group policy divergence index. Simultaneously, based on the temporal changes in edge weights in the resource coupling graph constructed in the first step, the degree of resource tension changes between different control cycles is calculated to obtain the group resource coupling tension index. Together, these two dual-index indicators characterize the current system operation status, reflecting behavioral deviation trends and revealing the level of resource conflict evolution, providing valuable insights into behavior-resource synergy assessment.

[0048] In step 3, the dual index indicators are fed into the trained game optimization prediction model. The model outputs a strategy behavior convergence risk score. If the score is below a first threshold, the strategy remains unchanged. If the score is above the first threshold, the refined analysis process proceeds to step 4. The dual index indicators obtained in step 2 are fed into the trained game optimization prediction model to predict whether the system has potential non-convergence risks based on the current strategy behavior state and resource coupling characteristics. This prediction model is a nonlinear structural model trained on historical multi-period evolution data. It can identify whether the complex game situation represented by the dual index will lead to drastic fluctuations in strategy behavior, local optimality traps, or imbalances in resource competition. The model numerically outputs a strategy behavior convergence risk score, which represents the stability of the multi-agent system's evolutionary path at the current stage. If the score is below the first threshold, the current system state is acceptable and no adjustments are required. If the score exceeds the first threshold, the system may face the risk of strategy divergence or increased resource conflict, requiring further refined evaluation.

[0049] Step 4: The policy trajectory and resource behavior of all intelligent agents are individually evaluated. Each agent's individual policy deviation index is calculated and compared with a second threshold. If the proportion of agents exceeding the second threshold relative to the total number of agents is lower than a set percentage, adjustments are made to the policy evolution direction and resource selection behavior of only those agents exceeding the threshold. If the percentage is higher than the set percentage, a unified policy update is performed for all agents. After entering the detailed analysis process, a more detailed analysis of the policy trajectory and resource behavior of all intelligent agents is performed. The system analyzes each agent's behavior trajectory individually, calculating its behavioral fluctuations and deviations over multiple control cycles to generate an individual policy deviation index. This index indicates the level of consistency between an individual agent and the overall behavioral pattern. The system compares each agent's individual policy deviation index with the second threshold, identifies the agents exceeding the threshold, and calculates the proportion of these agents relative to the total number of agents. If the ratio is lower than the preset ratio threshold, it means that the problem is mainly concentrated in a small number of individual behavioral deviations. The system will only make targeted adjustments to these agents that exceed the standard, including policy direction correction and resource selection priority compression. If the proportion of agents that exceed the standard is not lower than the set ratio, it means that the group behavior deviates significantly. The system will trigger a unified policy reconstruction process for all agents to achieve an overall recalibration of the system behavior.

[0050] Step 5: After the policy adjustment or unified update is complete, the edge weight parameters and agent behavior trajectory data in the resource coupling graph are updated. This serves as the basis for the next round of evaluation and optimization, forming a continuously iterative dynamic feedback loop to achieve adaptive convergence of policy evolution and global optimization of resource allocation behavior. After completing local adjustments for individual agents or unified policy updates for all agents, the resource coupling graph and agent behavior trajectory data described above need to be synchronously updated. This step is crucial for incorporating the results of adjustments within the current control cycle into the state initialization and evaluation calculations for the next cycle. Specifically, this involves adjusting edge weights in the resource coupling graph to reflect the coordinated or conflicting relationships between resources after the adjustments. It also involves updating agent behavior trajectories to record the behavioral output of each agent in the latest control cycle. This update mechanism forms a complete closed loop, enabling the system to make accurate decisions based on the latest state within each control cycle, achieving adaptive convergence in the policy evolution process, and driving resource allocation behavior toward a globally optimal and stable structure, ensuring the efficiency and stability of the entire multi-agent system over the long term.

[0051] In constructing a multi-dimensional, heterogeneous resource coupling map, we first categorize and classify the resource types involved in a multi-agent system, clearly defining them as at least computing resources, storage resources, communication resources, and energy resources. Computing resources refer to the processing power used to execute agent policy logic, such as CPU time slices; storage resources refer to the storage capacity used for data buffering or intermediate state storage; communication resources refer to the data transmission bandwidth used for interaction between agents or with external agents; and energy resources refer to the actual electricity and energy consumed during the execution of behaviors. These four resource categories constitute the system's core resource set and are heterogeneous and dynamic.

[0052] In each control cycle, the system collects the usage of the above four types of resources and obtains their usage intensity data. The usage intensity data refers to the frequency or occupancy ratio of each type of resource called by each intelligent agent per unit time. This data is input into the subsequent relationship calculation process as the original observation value. Using this original observation value, the synergy function and inhibition function between resource types are calculated in sequence. The synergy function is used to reflect the degree of positive mutual influence between two resource types in the actual agent task. Specifically, the higher the degree to which two resource types are called simultaneously in the same time period, the more synergistic they are in task execution, and the higher the value of the synergy function. For example, when a certain intelligent agent performs a video processing task, computing resources and communication resources are often called simultaneously, and the synergy function value of the two is relatively high. The synergy function can be calculated by the proportional relationship between the joint frequency of resource use and the frequency of their independent use.

[0053] The inhibition function reflects the negative constraint between two resource types in resource allocation. This refers to resource competition between two resources when there is a quota conflict, they cannot be allocated simultaneously, or they are mutually exclusive due to constraints. The inhibition function value increases. The inhibition function can be estimated by combining metrics such as the number of resource conflict events and the failure rate of proxy tasks. For example, if a computing resource is heavily used during a control cycle, resulting in the restricted use of an energy-consuming resource, this indicates a certain degree of inhibition between the two resources.

[0054] The results of the synergy function and the inhibition function are linearly weighted to form a composite weight. This linear weighted combination is not limited to a specific form; practitioners in the art can select and use it based on actual needs. This weight is used to express the comprehensive interaction between the two resource types in the current control cycle. The composite weight is directly used as the dynamic weight of the edge between resource type nodes in the graph structure, and this weight can be updated as the control cycle changes.

[0055] Specifically: in each control cycle, the number of resource allocation events in which each resource type participates in each intelligent agent behavior is counted, and the joint usage frequency of different resource types being called simultaneously in the same time period is calculated. The joint usage frequency is divided by the geometric mean of the individual usage frequency of each resource to obtain the initial value of the synergy function; by comparing the number of allocation conflict events between different resource types, that is, the number of times two resource types cannot be allocated simultaneously due to quota restrictions in resource allocation decisions, and combining the weights of task failure events caused by agent behaviors, the initial value of the inhibition function of one resource on another resource is calculated; the synergy function and the inhibition function are combined into the dynamic weight of the edge through normalized mapping, and the edge weight between resource types in the graph structure is updated to form the edge weight parameters of the tensor form graph that reflects the evolution of resource states.

[0056] The goal of constructing a graph structure with resource types as nodes and composite weights as edges is to establish a visual representation capable of representing the multidimensional interaction logic between resources. This structure allows the system to track the evolution of resource relationships in a graphical manner. To track resource status changes over time, the graph structure is further enhanced with time series identifiers. Time series identifiers refer to labeling each edge weight change with a corresponding control cycle number or timestamp, recording the evolution history of the graph structure over multiple consecutive control cycles. Ultimately, the resulting resource coupling graph, generated through this structuring process, is fed into subsequent evaluation and decision-making modules as a tensor-based graph, forming an important foundation for resource status modeling during the game optimization control process of multi-agent systems.

[0057] To calculate the swarm strategy divergence index, we first need to obtain the behavioral data of all intelligent agents executing their strategies over multiple consecutive control cycles, forming a strategy trajectory for each agent. A strategy trajectory is a record of the continuous decisions made by an intelligent agent based on the state of the environment over multiple control cycles, which can be represented as a time series of behavioral decisions.

[0058] To facilitate unified processing and analysis, each policy trajectory is time-series encoded. This means converting the original behavior sequence into a set of fixed-length numeric vectors in chronological order. Each vector represents the agent's policy behavior during a control cycle. For example, a vector can include dimensions such as policy type encoding, resource usage preferences, and current state feedback. After encoding all policy trajectories, a structured vector set representing the agent's behavior time series is formed.

[0059] During any control cycle, for each agent's policy vector, the Euclidean distance is calculated between it and the policy vectors of all other agents in the system during that cycle. Euclidean distance is the straight-line distance between two vectors in multidimensional space and is used to measure the differences in the behavioral characteristics of two agents. The Euclidean distances between the agent and all other agents are averaged to obtain the mean behavioral distance of the agent relative to the group during that cycle. A larger value indicates a greater degree of "dispersion" of the agent's behavior within the group.

[0060] The standard deviation of the agent's strategy vectors across the entire control cycle sequence is calculated, representing the internal volatility of its strategy behavior. The agent's divergence score is calculated using the agent's behavioral distance mean within a single cycle as the numerator and its cross-cycle standard deviation as the denominator. This score reflects the proportional relationship between the agent's behavioral deviation from the external group and the stability of its own behavior.

[0061] For example, if an agent's behavior fluctuates slightly from one cycle to the next, but differs significantly from that of other agents, its divergence score will be high, indicating that its behavior deviates from the group trend. Conversely, if the agent's behavior is both stable and consistent with the group's strategy, its divergence score will be low, indicating that its behavior is in the core convergence zone of the group.

[0062] The divergence scores of all intelligent agents in the system are collected and normalized. This involves uniformly scaling the scores so that they fall within the same dimensional range (e.g., between zero and one), eliminating the influence of varying scales. The arithmetic mean of all normalized divergence scores is then taken, assigning equal weight to each agent's contribution, resulting in the group strategy divergence index for the current control period. This group strategy divergence index measures the overall consistency of the strategic behavior of all intelligent agents within the current phase. Higher index values ​​indicate greater divergence between agent behaviors and a trend toward divergence in the system's strategy evolution. Lower index values ​​indicate more convergence in agent strategies and a more uniform and stable direction of system strategy evolution. This index serves as an important basis for determining whether adjustments to subsequent game optimization models are necessary.

[0063] When calculating the group resource coupling tension index based on the dynamic changes in edge weights within the resource coupling graph, it is first necessary to ensure that the structure of the resource coupling graph has been constructed within the control cycle. This graph consists of several resource-type nodes and the edges between them. Each edge has an edge weight that represents the strength of the relationship between resources. Edge weights are dynamic parameters that are continuously updated throughout the control cycle. To capture the degree of change in the strength of resource relationships across different cycles, the edge weights of all edges are extracted within two adjacent control cycles, and the weight change for each edge between these two cycles is calculated. The weight change is the edge weight in the current control cycle minus the edge weight in the previous control cycle, representing the fluctuation in the intensity of resource interactions.

[0064] The tension strength of each edge during this change is calculated as the square of the weight difference between the current cycle and the previous cycle, highlighting edges with large changes. This is multiplied by the edge's initial weight, which serves as a weight adjustment factor for its change magnitude, ultimately yielding the edge's tension strength. This process can be explained as edges with "high change + high initial coupling" generating a stronger tension response.

[0065] Taking each resource type node as the center, we traverse all edges connected to that node in the graph and summarize their corresponding tension strength values. This summary process calculates the total tension response between each resource type and other resource types within the current cycle, resulting in a local tension response value for that resource type. A larger local tension response value indicates a higher structural stress on that resource type within the current control cycle.

[0066] For example, if the edge weight between communication resources and energy resources fluctuates dramatically within a given cycle and the initial coupling weight is high, the corresponding tension intensity for that edge is high, and the local tension response values ​​at the communication and energy resource nodes increase accordingly, indicating that there is a significant cooperative fluctuation or competitive imbalance between these two resource types during that cycle. Finally, the local tension response values ​​for all resource types are weighted averaged to generate the final group resource coupling tension index. The weighting factor can be set based on the importance of each resource type in the system or its proportion of basic resources. This index measures the degree of system tension exhibited during the evolution of the overall resource state within the current control cycle. Higher index values ​​indicate more dramatic cooperative or conflicting fluctuations between resources within the cycle; lower values ​​indicate stable cooperative relationships between resources and a state of resource coordination and balance. The group resource coupling tension index, as a resource structural indicator within the dual-index evaluation system, is used in conjunction with the group strategy divergence index to determine whether subsequent strategy intervention and resource optimization actions are necessary.

[0067] When inputting the dual-index indicators into a trained game optimization prediction model and outputting a strategy behavior convergence risk score from the model, the group strategy divergence index and the group resource coupling tension index are first normalized. Normalization involves converting the original index values ​​to a uniform range (e.g., 0 to 1) to eliminate dimensional differences and dimension effects, ensuring that the two indices are equivalent in the model. This normalization method can use either the maximum-minimum normalization method or the Z-score normalization method to ensure comparability of index values ​​at different stages.

[0068] Based on the preset dimensionality requirements of the model's input structure, the two normalized indices are combined in a fixed order to form a dual-index input vector. This input vector represents the behavioral state of the current multi-agent system within the current control cycle, representing a joint characterization of the system's behavioral structure and resource coupling characteristics. This input vector is fed into a trained game optimization prediction model, a predictive network structure constructed based on data from multiple behavioral phases within the historical control cycle, and possesses nonlinear modeling capabilities. "Trained" means that the model has been trained and optimized on a sample set of historical behavioral data, forming a mapping relationship capable of predicting the risk of the input state.

[0069] This game optimization prediction model utilizes a nonlinear prediction architecture integrated with a feedforward neural network and a graph-associative reasoning structure. A feedforward neural network is a basic artificial neural network architecture suitable for learning the mapping between input vectors and output values, and can extract complex nonlinear features from static data. The graph-associative reasoning structure is used to process the portion of the input state that involves the evolution of the resource graph. Based on the historical trends in edge weights across graph structures, it can model the evolution of resource tension in the graph space, enhancing the model's ability to express the interactions between resources and behaviors.

[0070] The model searches for similar behavioral patterns within the historical behavior mapping space constructed in its training memory for the dual-index input vectors. Based on the positional relationships of the inputs within this space, it performs nonlinear fitting and inference prediction. Ultimately, the model outputs a continuous value representing the convergence risk score for the current input state during the policy evolution process. This score represents the likelihood that the entire multi-agent system is experiencing unstable evolution or potential divergence at the behavioral strategy level within the current control cycle. For example, if both the group strategy divergence index and the resource coupling tension index are in the medium-to-high range, indicating a relatively chaotic state of both behavior and resources, the model may output a high risk score. Conversely, if both indices remain low and fluctuate steadily, the model outputs a low risk score. This policy behavior convergence risk score serves as a basis for subsequent control decisions. If the score is below a preset first threshold, the system maintains the current policy. If it is above the threshold, the system proceeds to the next step of the detailed analysis process, individually evaluating each agent's behavior and triggering the policy intervention mechanism.

[0071] In the process of separately evaluating the policy trajectories and resource behaviors of all intelligent agents and calculating the individual policy deviation index of each agent, we first select the policy trajectory of each agent over multiple consecutive control cycles. The policy trajectory refers to the sequence of behavioral strategies executed sequentially by the agent in a continuous time series, which can reflect its behavioral evolution path.

[0072] This strategy trajectory is converted into a structured representation. Specifically, a strategy behavior vector is constructed for the agent within each control cycle. This vector contains key behavioral characteristics such as its strategy decision, resource selection preferences, and environmental state perception during that cycle. The strategy behavior vectors for all control cycles are arranged in chronological order to form a sequence of the agent's strategy behavior vectors.

[0073] A policy decision is the final action the agent selects from a set of optional behaviors within the current control cycle based on its policy function. For example, when faced with multiple tasks, it decides whether to execute a "high-energy, high-revenue task" or a "low-resource, conservative strategy." This feature can be written into a vector by mapping the different policy actions to discrete values. For example, the first dimension of the vector corresponding to "Task A" has a value of 1, while that of "Task B" is 2, and so on.

[0074] Resource selection preference refers to which resource types the agent prefers to use for behavior execution in the current control cycle. This preference can be expressed by statistically analyzing the order or weight of resource requests. For example, if the agent's resource requests in this cycle are ranked as "communication resources > storage resources > computing resources", a set of resource preference weight values ​​can be assigned to it in the vector, such as [0.5, 0.3, 0.2].

[0075] Environmental state perception refers to the agent's subjective reception and internal encoding of the environmental state before making strategic behaviors. For example, state indicators such as the current task complexity, neighboring agent density, and network communication delay are usually composed of the state feature encoding output by its perception module. For example, indicators such as high task complexity, 20% battery remaining, and weak communication signal can be standardized and used as vector dimension input, such as [0.8, 0.2, 0.1].

[0076] The three features described above—strategy decision, resource selection preference, and environmental state perception—are concatenated according to pre-set dimensions to form the agent's policy behavior vector for the current control cycle. For example, a vector containing a 3-dimensional policy decision code, a 3-dimensional resource preference value, and a 3-dimensional environmental state perception can be represented as a 9-dimensional vector: [1, 0, 0, 0.6, 0.3, 0.1, 0.7, 0.5, 0.2], where the first dimension represents the selected policy code, the second through fourth dimensions represent the resource preference weights, and the fifth through seventh dimensions represent the state perception parameters. The policy behavior vectors for all control cycles are arranged in chronological order to form the agent's policy behavior vector sequence, a two-dimensional structure of T × D, where T is the number of control cycles and D is the dimension of each behavior vector. This sequence captures the temporal dynamics of the agent's behavior evolution and is used in subsequent analyses to calculate behavioral volatility and group consistency. For example, if an agent generates five policy behavior vectors in five control cycles, each with the aforementioned 9-dimensional structure, its policy behavior vector sequence is a matrix with 5 rows and 9 columns.

[0077] Normalizing the vector sequence involves uniformly adjusting the values ​​of each dimension of each vector so that they are expressed on a comparable scale, eliminating interference caused by inconsistent dimensions or differences in the distribution of values ​​in the original data. Normalization can be performed using min-max scaling or zero-mean normalization.

[0078] Calculate the magnitude of the agent's policy behavior vector change between adjacent control cycles, that is, the degree of change between each two consecutive vectors. Average these magnitudes to obtain the agent's internal policy trajectory variability within the sequence. This metric indicates the agent's behavior's fluctuations over time. Larger values ​​indicate unstable behavior and frequent decision-making changes.

[0079] Select the strategy behavior vector sequences of other agents among all intelligent agents, align them to the same timeline and dimension as the current agent, and calculate the average vector difference between the current strategy behavior vector of the agent and the strategy behavior vectors of all other agents in the same control cycle as the degree of behavioral deviation between the agent and the group. This value represents the gap between the current decision of the agent and the group's strategy center.

[0080] The internal volatility of the strategy trajectory and the degree of behavioral deviation are weighted linearly combined—that is, they are weighted and added together according to two preset weighting factors—to form a single behavioral indicator for the agent in the current control cycle. This indicator combines the two dimensions of behavioral stability and group consistency to comprehensively evaluate the degree of abnormality in the agent's current behavior. To further eliminate the interference of occasional fluctuations, the behavioral indicator is smoothed within a fixed sliding window. The sliding window is a set number of consecutive control cycles. The indicator value is smoothed using a weighted moving average to highlight behavioral trends and mitigate temporary interference. This ultimately generates the agent's individual strategy deviation index for the current stage. A larger index value indicates that the agent's behavior is highly volatile and deviates significantly from the group, indicating a possible trend of behavioral abnormality. A smaller index value indicates that the agent's behavior is stable and consistent with the group's strategy, indicating that it is a convergent agent in the system. For example, if an agent's strategy has been changing continuously over the past five cycles and its strategy deviates significantly from the average behavior of other agents in most cycles, its individual strategy deviation index will be significantly higher than that of other agents, and the system will identify it as a potential source of instability. The individual strategy deviation index will be used in subsequent behavior judgments as a basis for deciding whether to modify the strategy evolution direction and adjust resource selection for the agent.

[0081] When adjusting the strategy evolution direction and resource selection behavior only for agents exceeding the second threshold, the system first obtains the strategy behavior vector of the agent exceeding the target during the current control cycle. This vector is a structured representation of the agent's behavior, encompassing multiple dimensions, including its strategy decision encoding, resource preference values, and environmental state perception parameters. Simultaneously, the system obtains the mean strategy behavior vector of all intelligent agents during the control cycle. This vector is then averaged across all dimensions to form the central representation of the group's behavior for the current cycle. By calculating the direction of the difference between the strategy behavior vector of the agent exceeding the target and the mean vector, a directional indicator of the deviation of the agent's current behavior from the group's evolutionary trend is obtained. Based on this difference direction, a behavior direction correction factor is set. This factor measures the degree of consistency between the deviation direction and a preset unit convergence vector. The unit convergence vector represents the standard evolution direction derived by modeling the convergence trend of the strategy trajectory during the training phase and is a normalized reference direction vector.

[0082] The behavior direction correction factor is calculated by calculating the cosine of the angle between the difference direction vector and the unit convergence vector (indicating the directional similarity between the two), taking the negative of this value (indicating that the correction direction should be adjusted in the opposite direction), and multiplying it by a non-zero scaling factor γ (gamma). γ is an adjustable control parameter that determines the magnitude of the adjustment. For example, if the angle is 60°, the cosine is 0.5, and γ is 0.8, the correction factor is −0.4. This value indicates that the agent's behavior has deviated from the group direction and should be adjusted back in the opposite direction by a certain amount in the next cycle. This correction factor is then added to the policy generation vector of the out-of-standard agent in the next control cycle. Specifically, the correction direction vector component is added to the behavior output stage of the policy generation model to fine-tune the behavior path, gradually aligning it with the group's convergence center.

[0083] The resource set used by the exceeding agent in the current control cycle, that is, the resource types actually applied for or occupied by it in the process of executing the strategy, is collected and compared with the resource usage frequency ranking of all agents in the current cycle. The resource types with relatively high usage frequency by the agent but low average usage frequency of the group are found, which are defined as the deviation resource set. These resource types represent the incoordinated resource preferences of the agent.

[0084] A compression factor η (eta) is set for the resource allocation priority of the outlier agent in the next control cycle. This factor is used to suppress the agent's probability of using deviant resources in its resource selection strategy. η is calculated as: the initial resource selection probability multiplied by [1-fd / (ft+C)], where fd is the agent's current frequency of use of a resource (e.g., the ratio of the number of times the resource is called to the total number of times the resource is available); ft is the group's average frequency of use of that resource; and C is a preset nonzero constant to avoid zero denominators and introduce inhibitory smoothing. This compression operation only applies to the resource types in the deviant resource set and does not affect the agent's freedom of action on the group's mainstream resources. Ultimately, the agent's resource policy function in the next cycle is dynamically adjusted to assign a lower probability to deviant resources, guiding its resource use behavior toward the group consensus preference, thereby achieving the goal of coordinated resource-level behavioral control.

[0085] Suppose agent A frequently uses "communication resources" and "energy resources" in the current cycle. However, according to group statistics, the frequency of "energy resources" usage is extremely low. Therefore, agent A's deviation from the group average in "energy resources" is significant. In this case, if A's usage frequency fd = 0.6, the group average ft = 0.2, and the constant C = 0.1, then the compression factor η = original adoption probability × [1 − 0.6 / (0.2 + 0.1)] = original adoption probability × [1 − 2.0] = original adoption probability × (−1.0). Since η cannot be negative, the system sets a lower limit of η to 0, meaning that this resource will be completely suppressed in the next cycle.

[0086] If the number of agents exceeding the second threshold exceeds a set percentage of all agents, the multi-agent system is experiencing widespread deviations in its policy behavior evolution, indicating a state of group-level convergence risk. In this case, the system will no longer focus solely on individual agent adjustments but instead trigger a unified policy update for all agents to restore behavioral consistency and improve resource allocation coordination. This update process first involves extracting the mean policy trajectory vector of all agents over the past N control cycles. The mean policy trajectory vector is a feature vector representing the long-term policy preference obtained by taking the weighted average of each agent's policy behavior vector over N control cycles. After calculating the mean trajectory vector for all agents, the resulting set is fed into a unified analysis model. This model further determines the average deviation of all agent behavior vectors relative to the mean trajectory vector. The average deviation measures the distance between an agent's behavior and the group consensus behavior and is typically calculated as the Euclidean distance between vectors. After calculating the average deviation of all agents, the strategic behavior vector with the smallest deviation is selected as the most representative behavior in the current system state. This vector is defined as the convergence direction vector, that is, the behavioral target direction that all agents should approach together in the future strategy evolution process.

[0087] Based on this convergence direction vector, the system resets the initial values ​​of all agent strategy generation parameters. Specifically, the convergence direction vector is multiplied by an initial scaling factor. This factor is a preset positive number that controls the amplitude and convergence speed of the newly generated strategy. A larger value indicates a more significant adjustment, while a smaller value indicates a smoother behavioral transition. This operation is equivalent to synchronously setting the initial conditions of each agent's strategy function to the weighted starting point in the consensus evolution direction, ensuring consistent behavioral direction.

[0088] The system also collaboratively reconstructs the agent's resource preferences. First, all resource behavior parameters are sorted by resource selection frequency, which represents the cumulative proportion of each resource type used by all agents over N control cycles. Based on this sorting result, primary and marginal resource sets are identified. Then, based on historical group behavior data, the co-occurrence probability of each resource type in the same strategy evolution phase is calculated—that is, the joint probability of a resource and other resources appearing simultaneously in agent behavior. Based on the co-occurrence probability matrix, a preferred resource interval is constructed between resource types. This recommends a set of resource combinations with high co-occurrence and low conflict for each strategy type. Using these resource intervals as a mapping reference structure, the agent's resource behavior parameters are uniformly remapped and reset, shifting their resource selection logic toward the preferred resource set and avoiding intensified system resource competition caused by individual preferences. In the new control cycle, all agents uniformly apply this synchronized behavior-resource mapping structure, using the convergence direction vector as the strategy direction constraint and the preferred resource interval as the resource selection boundary, thereby achieving synchronized convergence of strategy evolution and resource behavior.

[0089] Example 2: A multi-dimensional resource dynamic game optimization control system based on a multi-agent system, comprising:

[0090] The resource graph builder is used to construct a multi-dimensional heterogeneous resource coupling graph. By modeling the collaborative and inhibitory relationships between resource types within the same control cycle, it generates a graph structure with resource types as nodes and dynamic weights as edges, and outputs a tensor-based graph to describe the evolution of resource states.

[0091] A behavioral index calculator collects the strategy trajectories of all intelligent agents over multiple consecutive control cycles, calculates the group strategy divergence index and the group resource coupling tension index obtained by the dynamic changes of edge weights in the resource coupling graph, and generates a dual-index indicator to characterize the overall behavioral deviation degree and resource conflict status of the multi-agent system;

[0092] A game risk discriminator is used to receive the dual index indicator and input it into the trained game optimization prediction model, output the strategy behavior convergence risk score, and determine whether to trigger the subsequent analysis and adjustment process based on the comparison result of the score with the first threshold;

[0093] The local evolution controller is used to evaluate the strategy trajectory and resource behavior of all intelligent agents separately, calculate the individual strategy deviation index, and based on the comparison result of the index with the second threshold, determine whether the proportion of agents exceeding the standard is lower than the set ratio. If it is lower, only the strategy evolution direction and resource selection behavior of the exceeding agents are adjusted. If it is not lower, a unified strategy update operation is performed on all agents.

[0094] The evolutionary feedback updater is used to update the edge weight parameters and agent behavior trajectory data in the resource coupling graph after the strategy adjustment or unified update is completed, providing a new input basis for the next control cycle and forming an adaptive convergence feedback loop for strategy evolution.

[0095] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0096] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0097] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0098] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0099] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A multi-dimensional resource dynamic game optimization control method based on a multi-agent system, characterized by: The following steps are involved: Step 1: Construct a multi-dimensional heterogeneous resource coupling graph. By modeling the synergy and inhibition relationships between resource types within the same control cycle, a graph structure is formed with resource types as nodes and dynamic weights as edges, forming a tensor-form graph for resource state evolution. Resource types include computing resources, storage resources, communication resources, and energy resources. Step 2: Collect the strategy trajectories of all intelligent agents over multiple consecutive control cycles, calculate the group strategy divergence index and the group resource coupling tension index based on the dynamic changes of edge weights in the resource coupling graph, and use them as dual index indicators to characterize the overall behavior deviation and resource conflict status of the multi-agent system; Step 3: Input the dual index indicator into the trained game optimization prediction model. The model outputs a strategy behavior convergence risk score. If the score is below the first threshold, the strategy remains unchanged. If the score is above the first threshold, the detailed analysis process of step 4 is entered. Step 4: Evaluate the strategy trajectory and resource behavior of all intelligent agents individually, calculate the individual strategy deviation index of each agent, and compare this index with the second threshold. If the proportion of agents exceeding the second threshold among all agents is lower than the set ratio, adjust the strategy evolution direction and resource selection behavior only for the agents exceeding the second threshold; if the proportion is not lower than the set ratio, perform a unified strategy update operation for all agents; Step 5: After the strategy adjustment or unified update is completed, the edge weight parameters and agent behavior trajectory data in the resource coupling graph are updated. This serves as the basis for the next round of evaluation and optimization, forming a continuously iterative dynamic feedback loop to achieve adaptive convergence of strategy evolution and global optimization of resource allocation behavior.

2. The multi-dimensional resource dynamic game optimization control method based on a multi-agent system according to claim 1 is characterized in that: In the process of constructing a multi-dimensional heterogeneous resource coupling map, resource types are limited to at least computing resources, storage resources, communication resources, and energy consumption resources. In each control cycle, the usage intensity data of various resources are collected as raw observation values, and the synergy function and inhibition function between resource types are calculated using the raw observation values. The synergy function is used to reflect the degree of positive mutual influence between two resource types in actual tasks, and the inhibition function is used to reflect the negative constraint relationship between two resource types in allocation behavior; The calculation results of the cooperation function and the inhibition function are combined into a composite weight value, and the composite weight value is used as the dynamic weight of the edge to connect the corresponding resource type nodes, thereby constructing a graph structure with resource types as nodes and dynamic weights as edges. A timing mark is added to the graph structure to record the evolution of the structure in multiple consecutive control cycles.

3. The multi-dimensional resource dynamic game optimization control method based on a multi-agent system according to claim 2 is characterized in that: In the process of constructing a multi-dimensional heterogeneous resource coupling graph, the calculation method of the coordination function and inhibition function between resource types is based on the resource synchronization utilization rate and conflict occupancy rate within the control cycle, specifically: In each control cycle, the number of resource allocation events in which each resource type participates in each intelligent agent behavior is counted, and the joint usage frequency of different resource types being simultaneously called in the same time period is calculated. The initial value of the coordination function is obtained by dividing this joint usage frequency by the geometric mean of the individual usage frequencies of each resource. By comparing the number of allocation conflict events between different resource types, that is, the number of times two resource types cannot be allocated simultaneously due to quota restrictions in resource allocation decisions, and combining the weights of task failure events caused by agent behavior, the initial value of the inhibition function of one resource on another is calculated; The synergy function and the inhibition function are combined into the dynamic weight of the edge through normalized mapping, and the edge weight between resource types in the graph structure is updated to form the edge weight parameters of the tensor form graph that reflects the evolution of resource status.

4. The multi-dimensional resource dynamic game optimization control method based on a multi-agent system according to claim 3 is characterized in that: In the process of calculating the group strategy divergence index, the strategy trajectories of all intelligent agents in multiple consecutive control cycles are first encoded in time series, and each strategy trajectory is represented as a set of vectors of fixed length. Within the same control cycle, the mean Euclidean distance between each agent's strategy vector and the strategy vectors of all other agents is calculated, and the divergence score of the agent is obtained by the ratio of this mean to the standard deviation of the change of the agent's strategy vector itself. The divergence scores of all agents are then normalized and averaged together to obtain the group strategy divergence index. This index is used to measure the deviation amplitude, behavioral divergence, and consistency of evolutionary trends among all intelligent agents in their strategy behaviors.

5. The multi-dimensional resource dynamic game optimization control method based on a multi-agent system according to claim 4 is characterized in that: In the process of calculating the group resource coupling tension index based on the dynamic changes of edge weights in the resource coupling graph, the weight change values ​​of all edges are first extracted within two adjacent control cycles, and the tension intensity of each edge during the change process is calculated. The calculation method is to multiply the square of the weight difference between the current cycle and the previous cycle of each edge by the initial weight factor; then, with each resource type as the center, the tension intensity values ​​of all its connected edges are summarized to calculate the local tension response value of the resource type; then, the weighted average of the local tension response values ​​of all resource types is taken to generate the group resource coupling tension index, which is used to measure the degree of structural tension that occurs during the evolution of resource states within the current control cycle and the resource coordination stability at the system level.

6. The multi-dimensional resource dynamic game optimization control method based on a multi-agent system according to claim 5 is characterized in that: When inputting the dual-index indicators into the trained game optimization prediction model and having the model output the strategy behavior convergence risk score, the group strategy divergence index and the group resource coupling tension index are first normalized and mapped, and then a dual-index input vector is formed according to the preset dimensions. This input vector is input into the game optimization prediction model as an expression of the current multi-agent system behavior state; This game optimization prediction model is trained based on historical evolution data within the control cycle, and adopts a nonlinear prediction structure constructed by integrating a feedforward neural network and a graph association reasoning structure. By fitting and predicting the positional relationship of the input vector in the historical behavior mapping space, it outputs the convergence risk score of the current behavior state in the strategy evolution process.

7. The multi-dimensional resource dynamic game optimization control method based on a multi-agent system according to claim 6 is characterized in that: In the process of evaluating the strategy trajectories and resource behaviors of all intelligent agents separately and calculating the individual strategy deviation index of each agent, we first select the strategy trajectory of each agent in multiple consecutive control cycles and construct a sequence of its strategy behavior vectors in each control cycle; After normalizing the vector sequence, the change amplitude of the strategy behavior vector between adjacent control cycles is obtained, and the average change amplitude of the agent is used as the internal variability of its strategy trajectory; Select the strategy behavior vector sequence of other agents in all intelligent agents, and construct the strategy average vector difference of this agent relative to all other agents as the external difference of its behavior deviation degree; The internal difference between the strategy trajectory and the behavior deviation is combined into a single indicator in the form of a weighted linear combination, and the indicator is smoothed within a fixed window to finally generate the individual strategy deviation index of the agent at the current stage.

8. The multi-dimensional resource dynamic game optimization control method based on a multi-agent system according to claim 7 is characterized in that: When adjusting the strategy evolution direction and resource selection behavior only for the agent that exceeds the standard, i.e., the agent that exceeds the second threshold, a behavior direction correction factor is set based on the difference direction between the strategy behavior vector of the agent that exceeds the standard in the current control cycle and the strategy behavior mean vector of all intelligent agents. The correction factor is the negative of the cosine value of the angle between the difference vector and the unit convergence vector multiplied by a non-zero proportional coefficient γ. The correction factor is applied to the strategy generation vector superposition term in the next control cycle of the agent that exceeds the standard to adjust the direction of the strategy behavior evolution path; Collect the resource set used by the exceeding agent in the current control cycle, and form the deviation resource set with the resources with the largest difference from the current frequency ranking of all agent resources. Set the resource allocation priority compression factor η of the exceeding agent in the next cycle to be the initial resource selection probability multiplied by [1-fd / (ft+C)], where fd is the resource frequency used by the exceeding agent, ft is the group average resource usage frequency, and C is a preset non-zero constant. The compression operation is set to only act on the resource types in the deviation resource set.

9. The multi-dimensional resource dynamic game optimization control method based on a multi-agent system according to claim 8, characterized in that: If the number of agents exceeding the second threshold accounts for no less than a set proportion of all agents, a unified policy update operation is performed on all agents, including extracting the mean vector of the policy trajectories of all agents in the past N control cycles as a baseline behavior model, obtaining the average deviation of all agent behavior vectors in the model, and using the policy behavior vector with the smallest deviation as the convergence direction vector. The initial values ​​of the policy generation parameters of all agents are reset to the convergence direction vector multiplied by the initialization scale factor, and all resource behavior parameters are sorted by resource selection frequency and mapped to the preferred resource interval constructed based on the probability of group resource co-occurrence. The mapping structure is uniformly applied in the new control cycle.

10. A multi-dimensional resource dynamic game optimization control system based on a multi-agent system, based on the multi-dimensional resource dynamic game optimization control method based on a multi-agent system according to any one of claims 1 to 9, characterized in that: include: The resource graph builder is used to construct a multi-dimensional heterogeneous resource coupling graph. By modeling the collaborative and inhibitory relationships between resource types within the same control cycle, it generates a graph structure with resource types as nodes and dynamic weights as edges, and outputs a tensor-based graph to describe the evolution of resource states. A behavioral index calculator collects the strategy trajectories of all intelligent agents over multiple consecutive control cycles, calculates the group strategy divergence index and the group resource coupling tension index obtained by the dynamic changes of edge weights in the resource coupling graph, and generates a dual-index indicator to characterize the overall behavioral deviation degree and resource conflict status of the multi-agent system; A game risk discriminator is used to receive the dual index indicator and input it into the trained game optimization prediction model, output the strategy behavior convergence risk score, and determine whether to trigger the subsequent analysis and adjustment process based on the comparison result of the score with the first threshold; The local evolution controller is used to evaluate the strategy trajectory and resource behavior of all intelligent agents separately, calculate the individual strategy deviation index, and based on the comparison result of the index with the second threshold, determine whether the proportion of agents exceeding the standard is lower than the set ratio. If it is lower, only the strategy evolution direction and resource selection behavior of the exceeding agents are adjusted. If it is not lower, a unified strategy update operation is performed on all agents. The evolutionary feedback updater is used to update the edge weight parameters and agent behavior trajectory data in the resource coupling graph after the strategy adjustment or unified update is completed, providing a new input basis for the next control cycle and forming an adaptive convergence feedback loop for strategy evolution.

Images