Multi-agent asynchronous coordination method and system under centralized architecture

Abstract

The application discloses a kind of multi-agent asynchronous coordination method and system under centralized architecture, and is applicable to the task scheduling and resource coordination of heterogeneous intelligent agent in intelligent internet of things environment.The method uses main scheduling node centralized control mechanism, according to the real-time state information of multiple heterogeneous intelligent agents, combines with multi-agent game model to calculate scheduling priority function, realizes the asynchronous distribution and feedback control of task.Simultaneously introduce asynchronous synchronization window mechanism, still can keep scheduling continuity and high efficiency under the condition of incomplete information, and through agent feedback delay modeling and benefit function design realizes the fairness and robustness of task allocation.The application improves system throughput and response speed, and gives consideration to heterogeneity and scalability, and is applicable to multi-agent cooperation scene such as disaster emergency, industrial manufacturing, smart home, intelligent transportation.

Description

Technical Field

[0001] This invention relates to the field of multi-agent system coordination in the context of the Internet of Things (IoT), and particularly to a method and system for achieving asynchronous collaborative work of multiple heterogeneous intelligent agents under a centralized architecture. Background Technology

[0002] With the rapid development of IoT technology, an increasing number of heterogeneous devices (such as sensors, smart terminals, and edge computing nodes) are connecting to the network as intelligent agents to collaboratively complete complex tasks. However, in traditional centralized scheduling frameworks, a synchronous approach is typically used, waiting for feedback from all nodes before proceeding to the next scheduling step. This presents several challenges in heterogeneous IoT environments:

[0003] Heterogeneity Challenge: Different agents differ in computing power, power consumption, and network bandwidth, resulting in varying times required to complete the same task. If a unified scheduling mechanism is used, the overall system efficiency is limited by the slowest responding agent, severely reducing throughput and resource utilization.

[0004] The challenge of asynchronous coordination: In real-world IoT environments, intelligent agents may return results asynchronously due to network latency or their own processing time. How to coordinate asynchronously returned results and continuously advance the task process without forcing synchronization has become a key problem that needs to be solved in centralized control.

[0005] Lack of agent autonomy: In centralized scheduling frameworks, agent feedback is delayed, and the agent states acquired by the master scheduling node are also lagging, further exacerbating system instability and scheduling inefficiency. Therefore, it is urgent to introduce a master-slave game model to endow agents with autonomous decision-making capabilities under the control of the master node, coordinating asynchronous feedback and task execution to improve system stability and profitability.

[0006] Resource utilization and load balancing: In heterogeneous environments, high-performance agents may be repeatedly assigned tasks and become overloaded, while low-performance agents may be idle or slow to respond. How to fairly and effectively allocate tasks to balance the load and improve resource utilization requires new scheduling strategies.

[0007] Real-time performance and robustness: For tasks requiring real-time response, if some agents fail to provide feedback or malfunction, centralized scheduling needs to have a robust mechanism to continue running with incomplete information and make timely adjustments to avoid system blockage or performance degradation.

[0008] In summary, current technologies lack an efficient collaborative method that can balance the heterogeneity and autonomy of agents with the advantages of centralized scheduling and control. Therefore, there is an urgent need to propose a new scheduling mechanism and system architecture to achieve asynchronous collaboration and fair task allocation among multiple agents under centralized master node control, thereby improving the efficiency, real-time performance, and robustness of complex task processing in IoT environments. Summary of the Invention

[0009] This invention aims to overcome the problems of difficulty in coordinating asynchronous responses, uneven task allocation, low resource utilization and lack of robustness in the existing technology of multi-agent cooperative scheduling, and provides a multi-agent asynchronous cooperative method and system under a centralized architecture.

[0010] The method proposed in this invention adopts an architecture that combines centralized scheduling with autonomous decision-making by intelligent agents. A master scheduling node coordinates multiple heterogeneous intelligent agents to execute complex tasks. This method allows each intelligent agent to autonomously decide whether to execute a task based on its local state and reward function, and supports asynchronous feedback mechanisms. Thus, while improving the overall task processing throughput of the system, it ensures the autonomy and flexibility of individual intelligent agents.

[0011] A first aspect of the present invention provides a multi-agent asynchronous cooperative system under a centralized architecture, comprising:

[0012] The master scheduling node is used to receive task requests, obtain agent status, and perform scheduling priority calculation and task allocation.

[0013] Multiple heterogeneous intelligent agent nodes are used to receive task instructions, autonomously execute strategies, and asynchronously provide feedback on execution results.

[0014] The task management module is used to maintain the task queue and record task feedback.

[0015] The master scheduling node and each intelligent agent are connected through a network to form a star topology.

[0016] A second aspect of the present invention provides a multi-agent asynchronous cooperation method under a centralized architecture, applied to the above-mentioned system, comprising the following steps:

[0017] Step 1: The master scheduling node receives the tasks to be scheduled and obtains the parameter information of each agent.

[0018] Step 2: The master scheduler calculates the scheduling priority value for each candidate agent.

[0019] Step 3: Based on the priority calculation result, the main scheduling node assigns task instructions to the agent and waits for feedback asynchronously; after receiving the task instructions, the agent assigned the task calculates the reward function and determines whether to accept the task instructions.

[0020] Step 4: After completing the task or meeting the predetermined conditions, each agent reports the execution result or status information to the main scheduling node.

[0021] Step 5: The master scheduling node updates the state information of the corresponding agent based on the feedback and enters the next scheduling cycle.

[0022] Based on the above technical solution, the present invention has the following beneficial effects:

[0023] 1. Improve system throughput: This invention adopts an asynchronous collaborative mechanism, which allows the master scheduling node to enter the next scheduling cycle without waiting for all agents to complete synchronously. This effectively avoids the system from being blocked due to the slowest node, making task processing pipelined, thereby significantly improving the overall throughput efficiency of the system in complex task scenarios.

[0024] 2. Enhance the autonomous decision-making ability of intelligent agents: This invention allows intelligent agent nodes to autonomously determine whether to execute the task based on the local reward function after receiving the task instruction, giving each node a certain degree of policy autonomy, enhancing the autonomy and environmental adaptability of the multi-agent system, and is particularly suitable for complex and dynamically changing IoT application scenarios.

[0025] 3. Load balancing optimization: This invention introduces capability awareness and historical scheduling frequency, and dynamically adjusts the priority of intelligent agents in combination with fairness factors to avoid the problem of centralized task allocation, achieve balanced utilization of heterogeneous intelligent agent resources, and improve system stability and long-term operating efficiency.

[0026] 4. Reduced response latency: Through game theory-driven priority scheduling and asynchronous synchronization window mechanism, the global waiting time caused by individual low-speed intelligent agents is effectively reduced, thereby reducing the average system response latency, which shows advantages in IoT applications with high real-time requirements.

[0027] 5. Scalability and robustness: The master-slave architecture facilitates the expansion of the number of agents, and new agents can be seamlessly added to the schedule; asynchronous cooperation and robust scheduling strategies enable the system to tolerate the delay or failure of some agents, and maintain service continuity and performance stability even in the case of incomplete feedback.

[0028] In summary, this invention addresses the issues of collaborative latency, unfair scheduling, and system bottlenecks in heterogeneous intelligent agent environments by proposing a centralized asynchronous collaborative method that is highly efficient, fair, and robust. This method is suitable for task scheduling and resource management in large-scale intelligent IoT systems, demonstrating promising application prospects and engineering feasibility. Furthermore, this invention is also applicable to application scenarios requiring asynchronous collaborative task processing by multiple intelligent agents, such as intelligent transportation systems, smart home networks, intelligent manufacturing, and smart city infrastructure. Attached Figure Description

[0029] Figure 1 This is a functional diagram of a multi-agent asynchronous collaborative system under a centralized architecture according to an embodiment of this application, illustrating the information interaction relationship between the main scheduling node and multiple heterogeneous agents.

[0030] Figure 2 This is a task scheduling flowchart of a multi-agent asynchronous collaboration method under a centralized architecture according to an embodiment of this application, illustrating the process of the master node allocating tasks and the asynchronous execution and feedback of agents. Detailed Implementation

[0031] To better understand the present invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be emphasized that the following embodiments are intended to illustrate the technical principles and preferred implementations of the present invention, and are not intended to limit the scope of protection of the present invention. Based on the ideas of the present invention, those skilled in the art can make various modifications and substitutions in specific applications, and all such modifications and substitutions, as long as they do not deviate from the principles of the present invention, should fall within the scope of protection of the present invention.

[0032] See Figure 1 This application provides a multi-agent asynchronous collaborative system with a centralized architecture, including a master scheduling node, multiple distributed heterogeneous agents, and a task management module.

[0033] The master scheduling node (hereinafter referred to as the master node) serves as the central controller, used to receive new task requests from task sources or task queues, distribute task instructions to each agent via the network for execution, and collect feedback results from the agents.

[0034] The multiple distributed heterogeneous intelligent agents (Agent 1, Agent 2, ..., Agent N) are connected to the master node through a network. Each intelligent agent has different hardware performance and functional focus (such as computing intelligent agents, storage intelligent agents, sensing intelligent agents, etc.), and together they form a task execution agent pool.

[0035] The task management module is used to maintain the task queue and record task feedback.

[0036] The master scheduling node and each intelligent agent are connected through a network to form a star topology.

[0037] The information flow in the system is as follows: the master node retrieves tasks to be processed from the task queue, assigns the tasks to selected agents through the task instruction stream, and each agent autonomously executes the task and reports the results or status to the master node through the feedback stream. The master node updates the task queue and agent status based on the feedback and continuously schedules subsequent tasks. This architecture ensures that the master node has global control over the overall tasks, while allowing agents to work in parallel and asynchronously.

[0038] like Figure 2As shown in the figure, this application also provides a multi-agent asynchronous collaboration method under a centralized architecture, applied to the above-mentioned system. The method combines multi-agent game theory and a master-slave scheduling mechanism, and is applied to the collaborative task execution of multiple heterogeneous agents in a smart IoT system. It aims to solve problems such as asynchronous collaboration, resource allocation, and scheduling fairness in complex task processing. The process is briefly described as follows:

[0039] When a new task arrives, the master node first obtains the state information of all agents, including current load and available resources. Then, it calculates the priority or suitability of each agent to determine which agent should execute the task. If there are multiple tasks, the master node can also allocate them sequentially according to priority. Subsequently, the master node selects the agent with the highest priority and issues the task instruction. After receiving the task instruction, the agent will make an autonomous decision: based on its own state and reward function, it decides whether to accept the task. If the agent accepts the task, it begins asynchronous execution; if it refuses, it notifies the master node, which then selects a new agent. The master node does not need to block and wait, and can process other tasks in parallel or prepare for the next scheduling step. When the agent completes the task or reaches the predetermined feedback trigger condition, it sends the feedback result back to the master node. After receiving the agent's feedback, the master node updates the agent's state (e.g., releasing resource usage, recording completion time) and checks if there are any unprocessed tasks in the task queue. If there are new tasks, the master node will repeat the above scheduling process; if there are no tasks, it enters an idle waiting state.

[0040] Furthermore, the method specifically includes the following steps:

[0041] Step 1: The master scheduling node receives task requests from the task queue or external IoT environment, and obtains and senses the capability parameters, current availability status and expected feedback delay of each intelligent agent as the basis for subsequent scheduling decisions.

[0042] Step 2: Scheduling decision calculation: Based on the multi-agent game theory model, the master scheduling node calculates the scheduling priority value for each candidate agent by combining factors such as the computing performance, resource status, historical scheduling frequency and feedback delay of each agent; the priority calculation function can introduce a fair scheduling factor to avoid excessive concentration of tasks.

[0043] Step 3: Task Instruction Allocation and Autonomous Execution: Based on the priority calculation results, the master scheduler node allocates task instructions to one or more agents. Each agent receiving a task instruction autonomously evaluates the task's benefit based on its locally set benefit function and determines whether to accept the instruction. When the evaluation result meets preset acceptance conditions, the agent autonomously executes the corresponding task processing flow. If the evaluation result does not meet the acceptance conditions, the agent sends a rejection message to the master scheduler node so that the master scheduler node can make subsequent scheduling decisions.

[0044] Step 4: Asynchronous Result Feedback: After completing a task or meeting predetermined conditions, each agent reports the execution result or status information to the master scheduling node. The master node uses an asynchronous synchronous window mechanism, that is, it sets a time window or condition threshold and periodically collects feedback from agents. Even if it does not receive feedback from certain agents, it can still continue subsequent scheduling based on the information already collected.

[0045] Within the set asynchronous / synchronous window, when the feedback ratio exceeds the set threshold or the time exceeds the set threshold, the next round of scheduling is triggered, and tasks that do not provide feedback are considered to have timed out or been downgraded.

[0046] Step 5: State Update and Iteration: The master node updates the state information of the corresponding agent based on the received feedback, including task completion status, remaining resources, and historical behavior records. It also executes remedial strategies (such as reassigning or simplifying tasks) for tasks that time out or fail to respond. Then, it continues processing the next batch of task requests or enters the next scheduling cycle, iterating in this loop.

[0047] In particular, the method combines a multi-agent game theory model with an asynchronous cooperation mechanism. The key components are defined in a formula below:

[0048] (1) Scheduling Priority Function: To select a suitable executor among heterogeneous agents, the master node calculates a scheduling priority value for each agent. This priority function comprehensively considers factors such as the agent's capabilities, available resources, and historical performance. For example, agent i (Agent) can be defined... i Priority Pr for the current task i for:

[0049]

[0050] In the formula, C i Indicates Agent i A comprehensive capability index (such as a weighted combination of processing performance, bandwidth, etc.) i Indicates Agent i The current availability factor (e.g., the proportion of idle resources or the size of the idle time window), φ iD represents the fair scheduling factor. i This indicates that the task will be assigned to the Agent. i The expected delay or overhead (e.g., task queue waiting time, network latency, etc.), w c ,w a ,w a ε represents the corresponding weighting coefficient, and ε is a minimum constant to prevent division by zero.

[0051] The significance of the priority function lies in: the stronger the capability (C) i Larger), the more idle (A) i Larger), fewer historical scheduling times (φ) i The larger the expected delay (D) i Small) intelligent agents, their Pr i The larger the value, the better suited it is to undertake new tasks. The master node can be determined based on Pr. i The largest value is selected as the executing agent, thus ensuring efficiency while also considering heterogeneity (through C). i and A i Adjust the scores of different agents.

[0052] (2) Fairness factor: To avoid excessive concentration of tasks on certain agents and to ensure the fairness of the system, a fairness scheduling factor φ is introduced. i This is to adjust priorities and prevent excessive task concentration on a single agent. The fair scheduling factor φ... i Historical scheduling frequency f of the agent i Inversely proportional, for example:

[0053]

[0054] Where β is the adjustment coefficient, used to control the degree of influence of fairness on priority adjustment.

[0055] (3) Asynchronous Synchronous Window Mechanism: Considering the asynchronous feedback characteristics of each agent, this application introduces the concept of an asynchronous synchronous window in the scheduling process to balance waiting latency and continuous scheduling. The master node sets a time window length ΔT or a feedback ratio threshold. For example, it is stipulated that the master node waits a maximum of ΔT time to collect agent feedback in each round of scheduling, or it can proceed to the next scheduling step in advance when it receives more than a certain proportion (such as δ) of agent feedback. The synchronous window mechanism can be described by the following conditions:

[0056] If the following conditions are met:

[0057]

[0058] This triggers the scheduling mechanism, and scheduling continues.

[0059] The above formula indicates that the master node does not need to wait for feedback from all agents after each round of task allocation. It executes the scheduling decision for the next batch of tasks when it receives enough feedback (the proportion reaches δ) or the waiting time exceeds the threshold ΔT. Agents that fail to provide timely feedback will be considered to have timed out, and their results can be processed or remedied later.

[0060] The mechanism ensures partial system synchronization: it allows the results of most fast agents to be applied to the scheduler as early as possible, while preventing individual slow agents from indefinitely delaying the system's progress, thereby improving overall throughput and response speed.

[0061] (4) Agent Capability and Availability Modeling: To accurately evaluate heterogeneous agents, this embodiment models the capabilities and real-time availability of each agent. Based on the priority calculation formula, this embodiment models the agent capability C. i This can be broken down into combinations of multi-dimensional parameters. For example, let C... i =α1P i +α2M i +α3B i , where P i M i B i They represent Agents respectively i Standardized indicators such as processor performance, memory capacity, and communication bandwidth are used, with α1, α2, and α3 as corresponding weights, to integrate different resource dimensions into a single capability score.

[0062] Similarly, availability A i This can be modeled as the proportion of idle resources of an agent or a function reflecting the current load, for example: A i =1 / (1+L) i ), where L i Agent i The current number of queued tasks or CPU utilization. Through this modeling, the master node obtains the capability parameters and load status of each agent during scheduling, and substitutes them into the priority function to quantify and compare the relative advantages of each agent in executing candidate tasks.

[0063] (5) Scheduling Gain Game Modeling: This embodiment abstracts the task allocation problem into a multi-agent game model to balance agent autonomy and the global optimality of scheduling decisions. The game participants are defined as the master scheduling node and various agents, with the master node acting as the leader and the agents as followers, forming a master-slave game structure. For each task to be assigned, an Agent is defined. i The profit function U i It reflects the "benefits" and "costs" of an agent performing a task. For example:

[0064] Ui =R i -λ·K i

[0065] Where R i Indicates Agent i The reward value earned upon completing a task (which may be linked to the importance or size of the task), K i Indicates Agent i The cost incurred in performing this task (such as time and energy consumption, converted into a loss value) is represented by λ, which is a trade-off coefficient.

[0066] The payoff function measures the net utility of an agent performing a task. During the game, each agent strategically decides whether to pursue / accept a task based on its own state and payoff function (e.g., it might choose to reject a task request if its capabilities are insufficient or the cost is too high). The master node considers the payoffs of all agents and attempts to maximize global utility and scheduling fairness when allocating tasks. For example, global utility can be defined as follows:

[0067]

[0068] By selecting U total The optimal allocation scheme achieves Pareto-optimal task allocation results.

[0069] Since the master node possesses global information, it can employ a mechanism similar to auctions or bidding: each agent submits a bid for executing the task based on a reward function (implied in the priority Pr). i In the calculation, the master node, acting as the "auction host," selects the bidder with the highest and fairest benefit. Through this game theory modeling, when the system reaches Nash equilibrium, the strategies of each agent and the allocation decisions of the master node will tend to stabilize. This ensures that each agent rationally participates in cooperation based on its own circumstances, while also ensuring that the overall task allocation develops in the direction of optimizing system performance.

[0070] Within the above framework, this application's embodiments design the following key mechanisms to address specific problems in asynchronous collaboration:

[0071] (1) Agent Selection Strategy: The master node adopts an agent selection strategy that balances efficiency and fairness when allocating tasks. The priority function provides a preliminary ranking basis, but the master node can introduce a rotation factor or reward / penalty mechanism to avoid the situation where the same agent is always selected. For example, the priority of agents that frequently execute tasks recently can be appropriately reduced to accelerate their rest; the priority of agents that receive fewer tasks can be appropriately increased to enhance their participation. This strategy ensures that each agent has the opportunity to execute tasks, thereby preventing resource idleness and single-point overload, and achieving load balancing. In other words, agent selection depends not only on the best immediate performance, but also on historical scheduling frequency to achieve scheduling fairness.

[0072] (2) Feedback Delay Impact Modeling: To address the differences in task completion feedback times for intelligent agents, this embodiment explicitly models the impact of feedback delay on the system during scheduling decisions. The master node incorporates the expected delay D into priority calculation. i This policy lowers the priority of agents with slower expected responses, thus tending to assign tasks to agents with faster responses. Specifically, the master node can calculate the average response time for each agent and use this data to update the agent profile, allowing for more accurate predictions of D during future scheduling. i This enables more intelligent scheduling decisions.

[0073] (3) Robust scheduling strategy under incomplete feedback: In actual operation, some agents may be unable to provide timely feedback or even lose connection due to faults, network anomalies or other reasons. The embodiments of this application ensure that scheduling can continue without crashing even when feedback is incomplete through a robust strategy.

[0074] First, leveraging the asynchronous window and timeout mechanism, the master node performs fault-tolerant processing for tasks that fail to respond on time: this includes retrying sending commands, re-adding the task to the queue, or temporarily having other idle agents take over execution. When a partial agent failure occurs, the master node dynamically reconstructs the task allocation scheme based on the agent's capability model and the current system state to compensate for the missing agent's functionality. For example, if a critical agent of a certain category becomes unreachable, the master node can select another agent with slightly inferior capabilities but still capable of performing the task to complete the remaining tasks.

[0075] Secondly, the master node maintains a global state view of the entire system. Even if some information is missing, it can estimate the state based on historical data and feedback from other agents, thus adopting a degradation strategy in decision-making: for example, if the precise state of a certain agent is missing, it is assumed to be in a busy or inefficient state, and scheduling decisions are made in a more conservative manner. These mechanisms enable the system to adapt to uncertainty, ensuring that the task flow is not interrupted. In other words, in extreme cases, this method can degrade to reducing dependence on lagging agents and continue task execution by increasing scheduling frequency or expanding the strategic choice space, demonstrating high robustness.

[0076] In summary, this application achieves efficient asynchronous collaborative scheduling in complex IoT environments through a centralized master node + heterogeneous multi-agent architecture. The master node uses a game theory model to calculate task allocation strategies, making each round of scheduling a process of balancing global optimization and local rationality. The asynchronous synchronous window mechanism provides a trade-off between concurrent execution and finite waiting, improving system throughput and response speed. Simultaneously, fair selection and robust scheduling strategies are introduced to ensure balanced participation of all agents and the ability to handle abnormal situations. Whether in heterogeneous device collaboration in the Industrial Internet of Things, or in scenarios such as smart homes and drone swarm systems, this invention can provide a reliable multi-agent coordination solution.

[0077] It should be noted that the above embodiments are only used to illustrate the principle of the present invention. In practical applications, the algorithm parameters (such as weight w) can be adjusted according to specific needs. c w a w d The scheduling strategy can be adjusted and optimized by means of window thresholds ΔT, δ, etc., or it can be adaptively adjusted by combining machine learning methods, but these changes do not depart from the protection scope of this invention.

[0078] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-agent asynchronous cooperation method under a centralized architecture, applied to a multi-agent asynchronous cooperation system, the system comprising: The master scheduling node is used to receive task requests, obtain agent status, and perform scheduling priority calculation and task allocation. Multiple heterogeneous intelligent agent nodes are used to receive task instructions, autonomously execute strategies, and asynchronously provide feedback on execution results; The task management module is used to maintain the task queue and record task feedback; The main scheduling node and each intelligent agent are connected through a network to form a star topology. Its features include the following steps: Step 1: The master scheduling node receives the tasks to be scheduled and obtains the parameter information of each agent; Step 2: The master scheduler calculates the scheduling priority value for each candidate agent; Step 3: Based on the priority calculation result, the main scheduling node assigns task instructions to the agent and waits for feedback asynchronously; after receiving the task instructions, the agent assigned the task calculates the reward function and determines whether to accept the task instructions. Step 4: After completing the task or meeting the predetermined conditions, each agent reports the execution result or status information to the main scheduling node; Step 5: The master scheduling node updates the state information of the corresponding agent based on the feedback and enters the next scheduling cycle; Step 3 specifically involves: Based on the priority calculation results, the main scheduling node assigns task instructions to one or more agents; each agent receiving the task instruction autonomously evaluates the task reward based on its locally set reward function and determines whether to accept the task instruction accordingly. When the evaluation result meets the preset acceptance conditions, the agent autonomously executes the task and reports the execution status; when the evaluation result does not meet the acceptance conditions, the agent rejects the task and reports the current status or the reason for rejection to the main scheduling node so that the main scheduling node can make subsequent scheduling decisions. In step 4, the master scheduling node adopts an asynchronous synchronous window mechanism, specifically as follows: Set a time window or condition threshold to periodically collect feedback from the agent; within the set asynchronous / synchronous window, when the feedback ratio or time exceeds the set threshold, the next round of scheduling is triggered, and tasks that do not provide feedback are considered to have timed out or been downgraded. Step 5 further includes: In the event that the agent fails to provide feedback or malfunctions, the master scheduling node implements a robust scheduling strategy for incomplete feedback based on the historical average feedback time, the remaining queue length of the task, and the health status of the agent. The robust scheduling strategy is specifically as follows: First, with the help of asynchronous windows and timeout judgment, the master scheduling node performs fault tolerance processing for tasks that fail to respond on time; when a partial failure of an agent occurs, the master scheduling node dynamically reconstructs the task allocation scheme based on the agent's capability model and the current system state to replace the missing agent functions. Secondly, the master scheduling node maintains a global state view of the entire system. In the event of missing information, it makes estimates based on historical data and feedback from other agents, and adopts a degradation strategy when making decisions to ensure that the task flow is not interrupted.

2. The method according to claim 1, characterized in that, The parameter information of each agent includes the agent's capability parameters, current availability status, and expected feedback delay information.

3. The method according to claim 1 or 2, characterized in that, Step 2 specifically involves: The master scheduling node is based on a multi-agent game theory model, and calculates the scheduling priority function value for each candidate agent by taking into account the computing performance, resource status and feedback delay of each agent. The priority function provides a preliminary sorting basis. Based on this, the master scheduling node considers the historical scheduling frequency of each agent to avoid the situation where the same agent is always selected.

4. The method according to claim 3, characterized in that, The scheduling priority value is calculated as follows: The overall capability index of an intelligent agent is defined as a weighted combination of processing performance, memory capacity, and communication bandwidth; The availability coefficient is modeled as the proportion of idle resources of an agent or a function reflecting the current load; Set the corresponding parameter that represents the expected latency or overhead of assigning a task to an agent; The above indicators are weighted and combined, and a fair scheduling factor is introduced to adjust the priority and prevent the task from being overly concentrated on a single agent.

5. The method according to claim 1, characterized in that, The method for calculating the revenue function is as follows: The task allocation problem is abstracted into a multi-agent game model, with the game participants being the main scheduling node and each agent. The reward value obtained by the agent in completing the task and the cost incurred by the agent in performing the task are used to define the agent's reward function.

Images