A multi-agent dynamic task scheduling method based on game theory

Abstract

The application discloses a kind of multi-agent dynamic task arrangement methods based on game theory, by constructing the utility function of task based on potential game, the net utility value of each candidate task strategy executed by agent in effective game neighborhood set is calculated.Based on the minimum maximum regret criterion, the regret degree is calculated, and the regret degree is converted into strategy switching probability using the logic of Boltzmann distribution to generate a new task allocation scheme.The system performs distributed conflict resolution, compares the marginal utility value of each agent, determines the agent with the highest marginal utility value as the task executor, and triggers the suppression mechanism for the agent that does not obtain the execution right.The application ensures the consistency of individual decision and global target through potential game model, updates the strategy using regret degree to avoid local optimum, effectively solves the path conflict and resource competition problem in multi-agent cooperation, and improves the convergence speed and execution success rate of task arrangement.

Description

Technical Field

[0001] This invention belongs to the field of multi-agent cooperative control technology, and relates to a multi-agent dynamic task orchestration method based on game theory. Background Technology

[0002] In practical applications of multi-agent systems, such as drone swarms, autonomous vehicle fleets, or intelligent manufacturing automated guided vehicle clusters, agents often only acquire localized perception information. Existing task orchestration techniques typically rely on a central server for globally optimal solutions, which can lead to communication bandwidth bottlenecks and single-point-of-failure risks when the number of agents surges. Traditional distributed game theory methods usually assume that information interaction between agents is absolutely reliable and often only consider maximizing current gains. In dynamic environments, existing technologies lack consideration for the quality of historical interactions, easily leading to task allocation oscillations or deadlocks. Furthermore, to reach consensus, agents often need to engage in high-frequency, full-data interaction with all neighbors within their field of vision, resulting in unnecessary waste of computing power and energy.

[0003] Therefore, how to reduce communication overhead and achieve highly robust dynamic task orchestration in environments with weak communication and no global information center is a technical problem that urgently needs to be solved. Summary of the Invention

[0004] To address the problems existing in the background technology, this invention proposes a multi-agent dynamic task orchestration method based on game theory.

[0005] The first aspect of this invention provides a game theory-based multi-agent dynamic task orchestration method, comprising: Local state information of each agent in a multi-agent system and the set of neighboring agents; The credit score is calculated for each neighbor agent in the set of neighbor agents based on the spatiotemporal credit evaluation model, and the effective game neighbor set is selected according to the preset credit threshold. Construct a task utility function based on potential game theory to calculate the net utility value of the current agent in executing each candidate task strategy within the effective game neighbor set; The regret degree of the current strategy is calculated based on the minimum maximum regret criterion, and a probabilistic strategy update is performed based on the regret degree to generate a new task allocation scheme. Distributed conflict resolution is performed according to the new task allocation scheme to complete the dynamic task orchestration of multiple agents.

[0006] Optionally, the step of performing a credit score on each neighbor agent in the set of neighbor agents based on the spatiotemporal credit evaluation model, and selecting an effective game neighbor set according to a preset credit threshold, specifically includes: Obtain the difference between the current system time and the timestamp of the last successful communication between neighboring agents, as well as the real-time Euclidean distance between the current agent and its neighboring agents; The timeliness score is obtained by exponentially decaying the difference in timestamps using a time decay factor, and the spatial suitability score is obtained by calculating the deviation between the real-time Euclidean distance and the preset optimal communication distance using a distance sensitivity coefficient. The timeliness score and the spatial suitability score are linearly superimposed according to preset weights to generate a comprehensive credit score; If the overall credit score is greater than the preset credit threshold, the corresponding neighbor agent will be included in the effective game neighbor set.

[0007] Optionally, the construction of the task utility function based on potential game theory, which calculates the net utility value of the current agent executing each candidate task strategy within the effective game neighbor set, specifically includes: Determine the inherent priority value of the tasks to be executed as a positive benefit; The estimated energy cost of the agent moving to the task point and the waiting time cost caused by avoidance are calculated as negative deductions; Detect whether the current agent's strategy overlaps with the strategies of other agents in the effective game neighbor set in time and space. If so, generate a conflict penalty value. Subtracting the negative deduction and the conflict penalty from the positive benefit term yields the net utility value of implementing the strategy.

[0008] Optionally, the step of calculating the regret degree of the current strategy based on the minimum-maximum regret criterion, and performing probabilistic strategy updates based on the regret degree to generate a new task allocation scheme, specifically includes: Iterate through all feasible policies of the current agent and calculate the theoretical maximum net utility value of each feasible policy under the assumed state; The difference between the theoretical maximum net utility value and the actual net utility value corresponding to the current strategy is calculated to obtain the regret degree; Boltzmann distributed logic is used to convert the regret degree into a policy switching probability, wherein the regret degree is positively correlated with the policy switching probability. Based on the calculated strategy switching probability, a new task objective is selected from the feasible strategies for updating.

[0009] Optionally, the step of performing distributed conflict resolution according to the new task allocation scheme to complete the dynamic task orchestration of multiple agents specifically includes: Each agent broadcasts its own marginal utility value for the current task within a local scope; Within the effective game neighbor set, compare the marginal utility values ​​of each agent and determine the agent with the highest marginal utility value as the task executor; An agent that fails to obtain the right to execute a task triggers an inhibition mechanism, which stops bidding for the task within a preset time step.

[0010] Optionally, in the step of using the time decay factor to perform exponential decay calculation on the difference of the timestamps to obtain the timeliness score, the time decay factor is a value greater than 0 and less than 1, used to adjust the timeliness weight of historical interaction information.

[0011] Optionally, the logic for setting the conflict penalty value is as follows: if the current agent and the neighboring agent have selected the same task or planned path and there is spatiotemporal overlap, then the conflict penalty value is a preset maximum positive value; otherwise, it is zero.

[0012] Optionally, in the step of converting the regret degree into a strategy switching probability using Boltzmann distribution logic, a rationality coefficient is introduced to control the aggressiveness of strategy selection. When the rationality coefficient approaches infinity, it indicates a completely rational choice, and when it approaches zero, it indicates a random choice.

[0013] Optionally, the local state information includes the agent's current position coordinates, remaining battery power, and real-time perceived distribution of environmental obstacles.

[0014] Optionally, the method further includes: sending the task allocation instruction after dynamic task orchestration to the execution mechanism of the multi-agent system to drive the multi-agent system to perform the corresponding physical action.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a game theory-based method for dynamic task orchestration among multi-agent systems. By constructing a task utility function based on potential game theory, the individual payoffs of agents are tightly coupled with the global system objective, ensuring that the distributed decision-making process converges to Nash equilibrium, thus avoiding oscillations or deadlocks in dynamic environments. The use of the minimum-maximum regret criterion combined with Boltzmann distribution logic for policy updates allows agents to strike a balance between exploring new strategies and utilizing old experiences, effectively preventing the algorithm from getting trapped in local optima and improving the global optimization capability of task orchestration. By introducing a rationality coefficient, the system can flexibly adjust the aggressiveness of policy selection according to environmental changes, enhancing the adaptability of multi-agent collaboration. Furthermore, the combination of a distributed conflict resolution mechanism and a suppression mechanism based on marginal utility comparison ensures optimal resource allocation within a local scope, significantly reducing communication load and negotiation time between agents, and significantly improving the efficiency and robustness of multi-agent task execution in complex dynamic scenarios. Attached Figure Description

[0016] Figure 1This is a flowchart of a multi-agent dynamic task orchestration method based on game theory in one embodiment of the present invention; Figure 2 This is another flowchart of a game theory-based multi-agent dynamic task orchestration method in one embodiment of the present invention. Figure 3 This is another flowchart of a multi-agent dynamic task orchestration method based on game theory in one embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In one embodiment, such as Figure 1 As shown, a game theory-based multi-agent dynamic task orchestration method is provided, which is then applied to... Figure 1 Taking China as an example, the following specific steps will be used: S10: Obtain the local state information of each agent in the multi-agent system and the set of neighboring agents.

[0019] S20: Based on the spatiotemporal credit evaluation model, each neighbor agent in the set of neighbor agents is given a credit score, and an effective game neighbor set is selected according to a preset credit threshold.

[0020] Specifically, this invention acquires local state information through sensors mounted on each intelligent agent. This local state information includes the agent's current position coordinates in a Cartesian coordinate system, the remaining battery percentage, and the real-time distribution of environmental obstacles perceived by the lidar. The intelligent agent broadcasts this local state information to its surroundings via a wireless communication module and receives data from neighboring nodes, thereby constructing an initial set of neighboring intelligent agents.

[0021] Subsequently, to eliminate unreliable communication partners, the system assigns a credit score to each neighbor agent in the neighbor agent set based on a spatiotemporal credit assessment model. Specifically, the system obtains the difference between the current system time and the timestamp of the neighbor agent's last successful communication, and uses a time decay factor to exponentially decay this time difference to obtain a timeliness score. Simultaneously, the system calculates the real-time Euclidean distance between the current agent and its neighbor agents, and uses a distance sensitivity coefficient to calculate the deviation between this real-time Euclidean distance and a preset optimal communication distance, obtaining a spatial suitability score. The system then linearly superimposes the timeliness score and the spatial suitability score according to preset weights to generate a comprehensive credit score.

[0022] Finally, the system compares the comprehensive credit score with a preset credit threshold. If the comprehensive credit score is greater than the preset credit threshold, the neighboring agent is deemed to have reliable communication quality and is included in the effective game neighbor set. Through this processing method, the present invention can pre-filter out nodes with high communication latency or excessive distance before task orchestration, avoiding invalid information interfering with the decision-making process, thereby significantly reducing the complexity of subsequent game calculations and improving the collaborative efficiency and robustness of multi-agent systems in dynamic environments.

[0023] S30: Construct a task utility function based on potential game theory to calculate the net utility value of the current agent executing each candidate task strategy within the effective game neighbor set.

[0024] Specifically, this invention initializes a task utility function, which consists of a positive reward term, a negative deduction term, and a conflict penalty value. The system reads the inherent priority value of the task to be executed and uses it as the positive reward term. Subsequently, the system calculates the estimated energy cost for the agent to move to the task point based on the distance between the agent's current position and the task point, combined with the energy consumption coefficient per unit distance. Simultaneously, the system predicts the waiting time caused by avoidance based on the path planning algorithm, combined with the waiting time loss coefficient per unit time, and calculates the waiting time cost. The system adds the estimated energy cost and the waiting time cost together as the negative deduction term.

[0025] When calculating the conflict penalty, the system iterates through the candidate policies of other agents within the effective game's neighbor set. The system checks whether the current agent's policy overlaps with the policies of its neighbors in terms of time and space. If a planned path is detected occupying the same spatial region within the same time window, a conflict is determined, and the system uses a preset fixed value as the conflict penalty. If no conflict is detected, the conflict penalty is zero.

[0026] Finally, the system subtracts the sum of the negative deduction and conflict penalty from the positive reward to obtain the net utility value of implementing the strategy. Through this construction method, the present invention can transform the multi-agent cooperative problem into a potential game model. This model ensures that the process of individual agents pursuing the maximization of their own net utility value is consistent with the optimization direction of the system's global objective. This approach avoids agents getting trapped in local optima, effectively reduces path conflicts and resource competition among multiple agents when executing tasks, and improves the convergence speed and execution success rate of task orchestration.

[0027] S40: Calculate the regret degree of the current strategy based on the minimum maximum regret criterion, and perform probabilistic strategy update based on the regret degree to generate a new task allocation scheme.

[0028] Specifically, this invention first traverses all feasible strategies of the current agent within the effective game neighbor set. For each feasible strategy, the system calculates its theoretical maximum net utility value under the assumed state. Subsequently, the system obtains the actual net utility value corresponding to the strategy adopted by the current agent. The system calculates the difference between the theoretical maximum net utility value and the actual net utility value to obtain the regret degree of the current strategy.

[0029] Next, the system introduces a preset rationality coefficient to control the aggressiveness of strategy selection. Utilizing Boltzmann distribution logic, the system converts the calculated regret rate into a strategy switching probability. In this calculation, the regret rate and strategy switching probability have a positive correlation; that is, the higher the regret rate, the higher the corresponding strategy switching probability. Based on the calculated strategy switching probability, the system probabilistically samples all feasible strategies and selects a new task objective as the updated strategy, thereby generating a new task allocation scheme.

[0030] This invention utilizes the minimum-maximum regret criterion to guide the agent's decision-making updates, enabling the agent to quantify the potential losses resulting from not selecting the optimal strategy, thereby achieving a balance between exploring new strategies and utilizing old ones. This mechanism effectively avoids multi-agent systems from getting stuck in local optima or oscillating during dynamic task orchestration, significantly improving the system's adaptability in complex environments and the convergence stability of task allocation.

[0031] S50: Perform distributed conflict resolution according to the new task allocation scheme to complete the dynamic task orchestration of multiple agents.

[0032] Specifically, in this invention, each agent broadcasts its own marginal utility value for the current task via wireless communication within a local area. Subsequently, each agent receives and parses the broadcast information from other agents in the effective game neighbor set. The system compares the current agent's own marginal utility value with the marginal utility values ​​of its neighboring agents.

[0033] If the system determines that the current agent has the highest marginal utility value, it designates the current agent as the executor of the task. If the system determines that the marginal utility value of a neighboring agent is higher than that of the current agent, the current agent relinquishes its right to execute the task. In this case, the agent that does not obtain the task execution right triggers an inhibition mechanism. The system controls the agent to stop bidding for the task within a preset time step and either enters a dormant state or searches for other idle tasks.

[0034] Through the above processing method, this invention achieves a fully distributed conflict resolution mechanism. This method does not rely on global scheduling by a central controller; each agent can quickly resolve resource contention issues through local interactions. This mechanism effectively avoids path conflicts and resource deadlocks among multiple agents when executing tasks, significantly reduces communication bandwidth consumption, and improves the collaborative efficiency and real-time task orchestration of multi-agent systems in dynamic environments.

[0035] In one embodiment, such as Figure 2 As shown, in step S20, which involves scoring the credit of each neighbor agent in the set of neighbor agents based on the spatiotemporal credit evaluation model and selecting the effective game neighbor set according to a preset credit threshold, the following steps are also included: S21: Obtain the difference between the current system time and the timestamp of the last successful communication between the neighboring agent, as well as the real-time Euclidean distance between the current agent and the neighboring agent.

[0036] S22: The timeliness score is obtained by exponentially decaying the difference in timestamps using a time decay factor, and the spatial suitability score is obtained by calculating the deviation between the real-time Euclidean distance and the preset optimal communication distance using a distance sensitivity coefficient.

[0037] S23: The timeliness score and the spatial suitability score are linearly superimposed according to a preset weight to generate a comprehensive credit score.

[0038] S24: If the comprehensive credit score is greater than the preset credit threshold, then the corresponding neighbor agent will be included in the effective game neighbor set.

[0039] Specifically, this invention obtains the difference between the current system time and the timestamp of the last successful communication between neighboring agents. Simultaneously, the system calculates the real-time Euclidean distance between the current agent and its neighboring agents based on their location coordinates.

[0040] Next, the system uses a preset time decay factor to perform an exponential decay calculation on the difference in timestamps to obtain a timeliness score. This calculation results in lower scores for neighboring agents with greater communication delays. Subsequently, the system uses a distance sensitivity coefficient to calculate the deviation between the real-time Euclidean distance and the preset optimal communication distance to obtain a spatial suitability score. This calculation reduces the scores of neighboring agents that are too far or too close.

[0041] The system acquires preset timeliness and spatial weights, and linearly superimposes the timeliness score and spatial suitability score according to their corresponding weights to generate a comprehensive credit score. Finally, the system compares the comprehensive credit score with a preset credit threshold. If the comprehensive credit score is greater than the preset credit threshold, the communication quality of the neighboring agent is determined to be reliable, and it is included in the effective game neighbor set.

[0042] This invention quantifies the reliability of interactions between neighboring agents in both time and space. This method effectively eliminates nodes with high communication latency or unsuitable distances, preventing invalid information from interfering with subsequent game decisions. This mechanism significantly reduces communication overhead in multi-agent collaboration, improving the convergence speed of dynamic task orchestration and system robustness.

[0043] It is important to note that in the step of calculating the timeliness score by exponentially decaying the difference in timestamps using a time decay factor, the time decay factor is a value greater than 0 and less than 1, used to adjust the timeliness weight of historical interaction information. The time decay factor is set to a value greater than 0 and less than 1, for example, it can be 0.9 or 0.95. This parameter is mainly used to adjust the retention degree of historical interaction information in credit assessment. In specific calculations, the system obtains the difference between the current system time and the timestamp of the last successful communication with the neighboring agent to obtain the communication delay time. Subsequently, the system uses the time decay factor as the base and the communication delay time as the exponent to perform a power operation to generate a decay coefficient. Since the time decay factor is less than 1, the calculated decay coefficient will decrease exponentially as the communication delay time increases. The system uses this decay coefficient as the timeliness score, or multiplies it by a preset base score to obtain the timeliness score.

[0044] This invention can sensitively reflect real-time changes in the communication status of neighboring agents. When a neighboring agent has not communicated for a long time, its timeliness score will drop rapidly, thus quickly removing it from the effective game neighbor set. This mechanism ensures that agents only interact with recently active nodes with stable communication links, effectively avoiding task allocation conflicts caused by using outdated information, and significantly improving the collaborative reliability of multi-agent systems in dynamic environments.

[0045] In one embodiment, such as Figure 3 As shown, in step S30, which involves constructing the task utility function based on the potential game and calculating the net utility value of the current agent executing each candidate task strategy within the effective game neighbor set, the following steps are also included: S31: Determine the inherent priority value of the task to be executed as a positive benefit.

[0046] S32: Calculate the estimated energy cost of the agent moving to the task point and the waiting time cost caused by avoidance as a negative deduction.

[0047] S33: Detect whether there is spatiotemporal overlap between the current agent's policy and the policies of other agents in the effective game neighbor set. If so, generate a conflict penalty value.

[0048] S34: Subtract the negative deduction and the conflict penalty value from the positive benefit to obtain the net utility value of implementing the strategy.

[0049] Specifically, the system reads the task type identifier of the task to be executed, determines the inherent priority value of the task according to a preset task value table, and treats it as a positive reward item. Then, the system calculates the straight-line distance between the current agent's position and the task point coordinates, and calculates the estimated energy cost based on the energy consumption coefficient per unit distance. Simultaneously, the system predicts the waiting time caused by obstacle avoidance based on a path planning algorithm, and calculates the waiting time cost based on the time loss coefficient per unit time. The system adds the estimated energy cost and the waiting time cost together as a negative deduction item.

[0050] When calculating the conflict penalty, the system iterates through the candidate strategies of other agents within the effective game's neighbor set. The system checks if there is an overlap between the current agent's planned path and the planned paths of its neighbors on the timeline. If the same spatial region is detected within the same time window, a spatiotemporal conflict is determined, and the system uses a preset fixed value as the conflict penalty. If no conflict is detected, the conflict penalty is zero.

[0051] Finally, the system subtracts the sum of the negative deduction and conflict penalty from the positive reward to obtain the net utility value of implementing the strategy. Through this construction method, the present invention can transform the multi-agent cooperative problem into a potential game model. This model ensures that the process of individual agents pursuing the maximization of their own net utility value is consistent with the optimization direction of the system's global objective. This approach avoids agents getting trapped in local optima, effectively reduces path conflicts and resource competition among multiple agents when executing tasks, and improves the convergence speed and execution success rate of task orchestration.

[0052] It is worth noting that the logic for setting the conflict penalty value is as follows: if the current agent and the neighboring agent have selected the same task or planned the same path and there is a spatiotemporal overlap, then the conflict penalty value is a preset maximum positive value; otherwise, it is zero.

[0053] Specifically, this invention obtains the candidate strategies of the current agent and the candidate strategies of all neighboring agents in the effective game neighbor set. The system compares the target task identifiers of these strategies to determine whether the current agent and any neighboring agents have selected the same task to be executed. At the same time, the system compares the spatiotemporal trajectory data of each agent to determine whether there are overlapping areas in the time and space dimensions of the planned paths.

[0054] If the system detects that the current agent and a neighboring agent have selected the same task, or if the planned paths overlap in time and space, a resource contention or physical collision risk is identified. In this case, the system sets the conflict penalty value to a preset maximum positive value. This maximum positive value is configured to be significantly greater than the inherent priority value of the task. If no conflict is detected, the system sets the conflict penalty value to zero.

[0055] This invention utilizes significant numerical differences to construct a strong constraint mechanism in the utility function. When a conflict occurs, the large penalty value causes the net utility value of the strategy to drop sharply into negative territory. This approach forces agents to proactively avoid conflicting strategies when pursuing utility maximization, thereby eliminating potential collision risks and task duplication problems at the decision-making stage, significantly improving the smoothness and safety of multi-agent system collaborative operations.

[0056] In one embodiment, step S40, which involves calculating the regret degree of the current strategy based on the minimum-maximum regret criterion and performing a probabilistic strategy update based on the regret degree to generate a new task allocation scheme, further includes the following steps: S41: Traverse all feasible policies of the current agent and calculate the theoretical maximum net utility value of each feasible policy under the assumed state.

[0057] S42: Calculate the difference between the theoretical maximum net utility value and the actual net utility value corresponding to the current strategy to obtain the regret degree.

[0058] S43: The regret degree is converted into a policy switching probability using Boltzmann distribution logic, wherein the regret degree is positively correlated with the policy switching probability.

[0059] S44: Based on the calculated strategy switching probability, select a new task objective from the feasible strategies for updating.

[0060] Specifically, the system iterates through all feasible strategies of the current agent within its effective game neighbor set. For each feasible strategy, the system calculates its theoretical maximum net utility value under the assumed state. Then, the system obtains the actual net utility value corresponding to the strategy adopted by the current agent. The system calculates the difference between the theoretical maximum net utility value and the actual net utility value to obtain the regret degree of the current strategy. Next, the system introduces a preset rationality coefficient to control the aggressiveness of strategy selection. The system uses Boltzmann distribution logic to convert the calculated regret degree into a strategy switching probability. In this calculation process, the regret degree and the strategy switching probability have a positive mapping relationship; that is, the higher the regret degree, the higher the corresponding strategy switching probability. Based on the calculated strategy switching probability, the system probabilistically samples all feasible strategies and selects a new task objective as the updated strategy, thereby generating a new task allocation scheme.

[0061] This invention utilizes the minimum-maximum regret criterion to guide the decision-making updates of agents. This method enables agents to quantify the potential losses resulting from not selecting the optimal strategy, thereby achieving a balance between exploring new strategies and utilizing old ones. This mechanism effectively avoids multi-agent systems from getting trapped in local optima or oscillating during dynamic task orchestration, significantly improving the system's adaptability in complex environments and the convergence stability of task allocation.

[0062] It should be noted that in the step of converting the regret degree into the strategy switching probability using Boltzmann distribution logic, a rationality coefficient is introduced to control the aggressiveness of the strategy selection. When the rationality coefficient approaches infinity, it indicates a completely rational choice, and when it approaches zero, it indicates a random choice.

[0063] Specifically, when performing probability calculations, the system uses a preset rationality coefficient as an adjustment parameter. This rationality coefficient is used to control the agent's aggressiveness and level of certainty during policy selection.

[0064] In the specific computational logic, the system uses the product of the regret factor and the rationality coefficient of each feasible strategy as an exponential term, which is then substituted into the Boltzmann distribution function for calculation. When the rationality coefficient approaches infinity, the system exhibits a perfectly rational choice mode. In this mode, the differences in the exponential terms are greatly amplified, and the agent will choose the strategy with the highest regret factor with a probability close to 1, i.e., strictly choosing the optimal solution with the highest theoretical benefit. Conversely, when the rationality coefficient approaches zero, the system exhibits a random choice mode. In this mode, the values ​​of the exponential terms corresponding to each strategy tend to be consistent, and the probability of the agent choosing each feasible strategy tends to be equal. At this time, the agent's behavior is similar to Brownian motion, no longer solely dependent on the current utility evaluation result.

[0065] This invention enables flexible adjustment of the exploration and utilization behavior of a multi-agent system. In the early stages of task orchestration or during drastic environmental changes, adjusting the rationality coefficient towards zero encourages agents to conduct extensive exploration to discover potential better task allocation schemes. As the system stabilizes, adjusting the rationality coefficient towards infinity prompts agents to quickly converge to a Nash equilibrium point. This mechanism effectively avoids the system falling into local optima traps and significantly improves the global optimization capability and robustness of multi-agent collaboration in dynamic and complex environments.

[0066] In one embodiment, step S50, namely, performing distributed conflict resolution according to the new task allocation scheme to complete the dynamic task orchestration of multiple agents, further includes the following steps: S51: Each agent broadcasts its own marginal utility value for the current task within a local scope.

[0067] S52: Compare the marginal utility values ​​of each agent within the effective game neighbor set, and determine the agent with the highest marginal utility value as the task executor.

[0068] S53: An agent that fails to obtain the right to execute a task triggers an inhibition mechanism, which stops bidding for the task within a preset time step.

[0069] Specifically, in this invention, each agent broadcasts its marginal utility value for the currently pending task within its local communication range. This marginal utility value reflects the net increase in benefit for the agent in performing the task relative to its current state. Subsequently, the system performs a numerical comparison within its effective game neighbor set, selects the agent with the highest marginal utility value, and identifies it as the sole executor of the task.

[0070] For agents that fail to secure task execution rights, the system triggers a suppression mechanism. This mechanism forces unsuccessful agents to cease bidding for the task within a preset time step. During this period, unsuccessful agents will either re-evaluate the remaining unassigned tasks or wait for the next decision cycle.

[0071] This invention utilizes the comparison of marginal utility values ​​to achieve distributed task allocation. This mechanism ensures that, locally, tasks are always assigned to the agent that generates the maximum system gain, thereby improving the overall efficiency of task orchestration. Simultaneously, the suppression mechanism effectively avoids repeated bidding and oscillations among multiple agents for the same task, significantly reducing the load on communication links and accelerating the convergence speed of the multi-agent system globally.

[0072] Optionally, the game theory-based multi-agent dynamic task orchestration method further includes: sending the task allocation instruction after dynamic task orchestration is completed to the execution mechanism of the multi-agent system to drive the multi-agent system to perform the corresponding physical action.

[0073] Specifically, after the game theory-based multi-agent dynamic task orchestration method completes calculations and determines the final task allocation scheme, the system generates task allocation instructions containing target coordinates, action types, and execution sequences. The system then sends these task allocation instructions to the underlying execution mechanisms of the corresponding agents in the multi-agent system via a communication interface.

[0074] After receiving the instruction, the actuator parses the control parameters within it. The actuator then invokes a motor driver or hydraulic control unit to drive the intelligent agent's mobile chassis or robotic arm to perform corresponding physical actions. For example, the intelligent agent may move to a designated work point according to the instruction, or operate an end effector to complete target grasping and placement.

[0075] This invention achieves closed-loop control from the decision-making layer to the execution layer. This mechanism ensures that the optimization strategies generated by the multi-agent system in a dynamic environment can be transformed into actual physical behaviors, effectively avoiding the problem of disconnect between decision-making and execution, and significantly improving the task execution accuracy and automated operation efficiency of the multi-agent system in complex physical scenarios.

[0076] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0077] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-agent dynamic task orchestration method based on game theory, characterized in that, include: Obtain the local state information of each agent in a multi-agent system and the set of neighboring agents; The credit score is calculated for each neighbor agent in the set of neighbor agents based on the spatiotemporal credit evaluation model, and the effective game neighbor set is selected according to the preset credit threshold. Construct a task utility function based on potential game theory to calculate the net utility value of the current agent in executing each candidate task strategy within the effective game neighbor set; The regret degree of the current strategy is calculated based on the minimum maximum regret criterion, and a probabilistic strategy update is performed based on the regret degree to generate a new task allocation scheme. Distributed conflict resolution is performed according to the new task allocation scheme to complete the dynamic task orchestration of multiple agents.

2. The game theory-based multi-agent dynamic task orchestration method according to claim 1, characterized in that, The method of using a spatiotemporal credit evaluation model to assign a credit score to each neighbor agent in the set of neighbor agents, and selecting an effective game neighbor set based on a preset credit threshold, specifically includes: Obtain the difference between the current system time and the timestamp of the last successful communication between neighboring agents, as well as the real-time Euclidean distance between the current agent and its neighboring agents; The timeliness score is obtained by exponentially decaying the difference in timestamps using a time decay factor, and the spatial suitability score is obtained by calculating the deviation between the real-time Euclidean distance and the preset optimal communication distance using a distance sensitivity coefficient. The timeliness score and the spatial suitability score are linearly superimposed according to preset weights to generate a comprehensive credit score; If the overall credit score is greater than the preset credit threshold, the corresponding neighbor agent will be included in the effective game neighbor set.

3. The game theory-based multi-agent dynamic task orchestration method according to claim 1, characterized in that, The construction of the task utility function based on potential game theory, which calculates the net utility value of the current agent executing each candidate task strategy within the effective game neighbor set, specifically includes: Determine the inherent priority value of the tasks to be executed as a positive benefit; The estimated energy cost of the agent moving to the task point and the waiting time cost caused by avoidance are calculated as negative deductions; Detect whether the current agent's strategy overlaps with the strategies of other agents in the effective game neighbor set in time and space. If so, generate a conflict penalty value. Subtracting the negative deduction and the conflict penalty from the positive benefit term yields the net utility value of implementing the strategy.

4. The game theory-based multi-agent dynamic task orchestration method according to claim 1, characterized in that, The process of calculating the regret degree of the current strategy based on the minimum-maximum regret criterion, and then performing probabilistic strategy updates based on the regret degree to generate a new task allocation scheme, specifically includes: Iterate through all feasible policies of the current agent and calculate the theoretical maximum net utility value of each feasible policy under the assumed state; The difference between the theoretical maximum net utility value and the actual net utility value corresponding to the current strategy is calculated to obtain the regret degree; Boltzmann distributed logic is used to convert the regret degree into a policy switching probability, wherein the regret degree is positively correlated with the policy switching probability. Based on the calculated strategy switching probability, a new task objective is selected from the feasible strategies for updating.

5. The game theory-based multi-agent dynamic task orchestration method according to claim 1, characterized in that, The step of performing distributed conflict resolution according to the new task allocation scheme to complete the dynamic task orchestration of multiple agents specifically includes: Each agent broadcasts its own marginal utility value for the current task within a local scope; Within the effective game neighbor set, compare the marginal utility values ​​of each agent and determine the agent with the highest marginal utility value as the task executor; An agent that fails to obtain the right to execute a task triggers an inhibition mechanism, which stops bidding for the task within a preset time step.

6. The game theory-based multi-agent dynamic task orchestration method according to claim 2, characterized in that, In the step of using the time decay factor to perform exponential decay calculation on the difference of the timestamps to obtain the timeliness score, the time decay factor is a value greater than 0 and less than 1, which is used to adjust the timeliness weight of historical interaction information.

7. The game theory-based multi-agent dynamic task orchestration method according to claim 3, characterized in that, The logic for setting the conflict penalty value is as follows: if the current agent and the neighboring agent have selected the same task or planned path and there is spatiotemporal overlap, the conflict penalty value is a preset maximum positive value; otherwise, it is zero.

8. The game theory-based multi-agent dynamic task orchestration method according to claim 4, characterized in that, In the step of converting the regret degree into the strategy switching probability using Boltzmann distribution logic, a rationality coefficient is introduced to control the aggressiveness of strategy selection. When the rationality coefficient approaches infinity, it indicates a completely rational choice, and when it approaches zero, it indicates a random choice.

9. The game theory-based multi-agent dynamic task orchestration method according to claim 1, characterized in that, The local state information includes the agent's current location coordinates, remaining battery power, and real-time perceived distribution of environmental obstacles.

10. The game theory-based multi-agent dynamic task orchestration method according to claim 1, characterized in that, The method further includes sending the task allocation instruction after dynamic task orchestration is completed to the execution mechanism of the multi-agent system to drive the multi-agent system to perform the corresponding physical action.

Images