A multi-agent task scheduling method and system based on a distributed architecture

Abstract

This invention belongs to the field of distributed systems and artificial intelligence technology, and relates to a multi-agent task scheduling method and system based on a distributed architecture, applicable to heterogeneous task collaboration and dynamic scheduling in scenarios such as game theory and adversarial competition. The method includes: dividing the original task into sub-task sets through multi-dimensional adaptive granularity; evaluating the capabilities of the agents, and scheduling each sub-task in the sub-task set to multiple agents for execution based on the agents' capabilities; and aggregating the sub-task execution results of each agent to obtain the final task execution result. This invention can solve the problems of single point of failure, poor scalability, insufficient resource awareness, and task conflicts existing in traditional centralized scheduling systems. By constructing a "decentralized + self-organizing + collaborative decision-making" agent network, and utilizing a lightweight consensus mechanism and dynamic task splitting strategy, it achieves distributed, dynamic, and adaptive task scheduling among multiple agents.

Description

Technical Field

[0001] This invention relates to the fields of distributed systems and artificial intelligence technology, specifically to a multi-agent task scheduling method and system based on a distributed architecture, applicable to heterogeneous task collaboration and dynamic scheduling in scenarios such as game competition. Background Technology

[0002] This invention primarily utilizes a distributed architecture and a situational awareness center to address the problems of multi-agent scheduling and resource coordination. In the field of multi-agent systems and distributed task scheduling, there exists a wealth of widely recognized and applied technological foundations, which form the "technological soil" upon which this invention is built and innovated. The following are the well-known and generally accepted technological foundations in this field: (1) Foundations of Distributed Architecture. Distributed system architecture is a common architecture for modern intelligent systems (such as the Internet of Things, edge computing, and robot swarms), and its core characteristics are "decentralization, modularity, and scalability". Well-known technologies include: microservice architecture, event-driven architecture, message bus, service discovery mechanism, etc.

[0003] (2) Foundations of Multi-Agent Collaboration. Multi-agent systems (MAS) are a mature paradigm in artificial intelligence and control engineering, and are widely used in fields such as robot collaboration and intelligent transportation. Known technologies include: inter-agent communication protocols, behavior planning models (such as rule-based decision-making and reinforcement learning strategies), and task decomposition and allocation algorithms (such as A*, Dijkstra's algorithm, and genetic algorithm).

[0004] (3) Basic Task Scheduling Mechanism. Task scheduling is a core function of automated systems. Well-known technologies include priority scheduling, time-slice round-robin, resource preemption mechanism, and graph-based dependency modeling.

[0005] (4) Situational awareness and information fusion. Perception and fusion are the foundation of an intelligent agent's "cognitive environment". Well-known technologies include: Kalman filtering, particle filtering, Bayesian networks, target tracking algorithms (such as SORT, DeepSORT), etc.

[0006] (5) Resource management and allocation basis. The allocation of resources (such as computing, communication bandwidth, energy, and physical space) is a key issue in multi-agent systems. Well-known technologies include: auction mechanisms, game theory-based allocation, greedy strategies, and load balancing algorithms.

[0007] Traditional task scheduling methods are usually based on a centralized architecture, which has the following problems: (1) High risk of single point of failure: All scheduling decisions are concentrated in a central node. Once the node crashes or the network is interrupted, the entire scheduling system will be paralyzed; (2) Poor scalability: When the number of agents or the scale of tasks increases, the processing capacity of the central node will quickly become saturated, resulting in increased latency and slower response; (3) Weak resource awareness: Traditional methods lack dynamic awareness and adaptive adjustment of the real-time status of each agent (such as load, energy consumption, and bandwidth); (4) Insufficient task conflict handling capability: There may be task overlap or resource competition among multiple agents, and there is a lack of conflict detection and negotiation mechanisms. In addition, existing distributed task scheduling schemes mostly adopt static rules or simple communication mechanisms, which lack the ability to coordinate and optimize task granularity, execution path, and agent capabilities.

[0008] Therefore, there is an urgent need for a multi-agent task scheduling method that takes into account high availability, scalability, dynamic adaptability and task collaboration. Summary of the Invention

[0009] The purpose of this invention is to provide a multi-agent task scheduling method and system based on a distributed architecture, which solves the problems of single point of failure, poor scalability, insufficient resource awareness, and task conflicts in traditional centralized scheduling systems, and realizes efficient, dynamic, and adaptive scheduling of tasks among heterogeneous agents.

[0010] The technical solution adopted in this invention is as follows: A multi-agent task scheduling method based on a distributed architecture includes the following steps: The original task is divided into subtasks by multi-dimensional adaptive granularity. Evaluate the capabilities of the agents and, based on those capabilities, schedule each subtask in the subtask set to be executed by multiple agents. The results of the subtasks performed by each agent are aggregated to obtain the final task execution result.

[0011] Furthermore, the original task is divided into multi-dimensional adaptive granularities, including vertical decomposition based on business logic, horizontal slicing based on computational complexity, and spatiotemporal alignment based on execution time.

[0012] Furthermore, the evaluation of the agent's capabilities includes: using a resource matching degree function to evaluate the agent's capabilities based on the resources required for the task and the current resource state vector.

[0013] Furthermore, the process of scheduling each subtask in the subtask set to multiple agents for execution includes time-triggered scheduling and event-triggered scheduling. The time-triggered scheduling uses a comprehensive scoring mechanism to allocate tasks based on the time of task scheduling, quantifying and integrating three key indicators: resource matching degree, distance affinity, and load idle degree, to achieve optimal agent selection. The event-triggered scheduling drives the start, stop, execution, or state change of tasks through specific events triggered by agents.

[0014] Furthermore, the global situation map is constructed, maintained, and distributed using the situation center. This global situation map provides global information to each agent, enabling each agent to maintain real-time awareness of resources and the environment during the execution of sub-tasks.

[0015] Furthermore, a distributed architecture is adopted to deploy a master scheduling center and slave centers, which are used to schedule each subtask to multiple intelligent agents for execution. The master scheduling center is the top-level decision-making and macro-allocation node, while the slave centers are local scheduling and decision-making nodes. The slave centers receive task packages issued by the master scheduling center and allocate the subtasks to intelligent agents.

[0016] Furthermore, after all subtasks are completed, the main scheduling center or a designated agent aggregates the execution results of each subtask and triggers a feedback mechanism to optimize the next round of task scheduling.

[0017] A multi-agent task scheduling system based on a distributed architecture, comprising: The task partitioning module is used to divide the original task into subtask sets by multi-dimensional adaptive granularity. The task scheduling module is used to evaluate the capabilities of the agents and schedule each subtask in the subtask set to be executed by multiple agents based on the capabilities of the agents. The aggregation module is used to aggregate the subtask execution results of each agent to obtain the final task execution result.

[0018] The beneficial effects of this invention are as follows: This invention proposes a multi-agent task scheduling method based on a distributed architecture. Its core idea is to construct a decentralized + self-organizing + collaborative decision-making agent network, utilizing a lightweight consensus mechanism and a dynamic task splitting strategy to achieve distributed, dynamic, and adaptive task scheduling among multiple agents. The lightweight consensus mechanism in this invention is manifested as a "situation-driven, rule-based synchronization" strategy, using a situation center as the aggregation point for global data to ensure that all agents acquire a consistent global perspective within a very short time.

[0019] Decentralization in this invention refers to no longer relying on a single supercomputer to control all details. The master center only performs macro-level distribution, while slave centers and workers have local decision-making power. Through a master-slave distributed scheduling design, the global scheduling pressure is distributed. The master center is responsible for task registration and macro-level allocation, while the decision-making power for task start / stop and resource preemption is decentralized to slave centers. Even if the master center fails, slave centers and workers can still maintain local operation (such as event triggering) through local caching and event broadcasting.

[0020] The self-organization in this invention refers to the intelligent agent automatically adjusting its behavior based on real-time status and environmental events without waiting for instructions. When a node fails, other nodes automatically fill the gap based on "resource matching degree"; task granularity is automatically split and reorganized as the environment changes, without the need for manual intervention.

[0021] The collaborative decision-making in this invention refers to a closed-loop collaborative decision-making system driven by a "situation center" to address information asymmetry and decision-making conflicts in a distributed environment. The situation center acts as the "glue" for collaboration, aggregating all reported state and environmental data in real time, ensuring that each agent makes decisions based not only on local information but also on a global perspective. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the implementation principle and architecture of the multi-agent task scheduling method based on a distributed architecture of the present invention.

[0023] Figure 2 This is a flowchart of the main steps of the multi-agent task scheduling method based on a distributed architecture of the present invention.

[0024] Figure 3 This is an event scheduling flowchart in the multi-agent task scheduling method based on a distributed architecture of the present invention.

[0025] Figure 4 This is a block diagram of the multi-agent task scheduling system based on a distributed architecture according to the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0027] The main contents and key points of the multi-agent task scheduling method based on a distributed architecture of the present invention include: (1) Task modeling and granular division. The original task is divided into multiple executable subtasks according to the dimensions of business logic, computational complexity, execution time, etc., and then the tasks are registered.

[0028] (2) Agent capability assessment. Each agent reports its operational status in real time, forming a "capability profile".

[0029] (3) Task allocation and scheduling. Task scheduling is divided into time-triggered scheduling and event scheduling to achieve dynamic response and intelligent collaboration of tasks. In the scheduling process, indicators such as resource matching degree, distance and load are considered.

[0030] (4) Task Execution and State Synchronization. Each subtask is executed by an agent. To address issues such as resource conflicts, information asymmetry, and inconsistent decision-making among agents, a "situation center" mechanism is introduced. During agent execution, resources and the environment are kept real-time to ensure consistency between resource consumption and state. The agent reports its state (such as progress, errors, and resource consumption) in real time during execution and synchronizes it to the system server via a lightweight broadcast mechanism.

[0031] (5) Result aggregation and task closure. After all subtasks are completed, the results are aggregated by the main scheduling center or a designated agent to form the final task output and trigger a feedback mechanism to optimize the next round of scheduling.

[0032] The implementation principle and system architecture of the multi-agent task scheduling method based on distributed architecture of the present invention are as follows: Figure 1 As shown in the diagram. System services refer to the system server, which is the backend service connecting the front-end interactive interface and the underlying intelligent agents in this system. As the data hub and communication gateway of the distributed architecture, it undertakes the core responsibility of bidirectional data flow. The main scheduling center is the top-level decision-making and macro-allocation node of the scheduling center, used for task registration and initialization, macro-load balancing, and policy distribution. The slave centers are local scheduling and decision-making nodes, performing local task distribution: they receive task packages from the main scheduling center and assign specific sub-tasks to appropriate workers. Workers are specific intelligent agents that independently execute specific sub-tasks. The situation center is the core logical module in this system responsible for building, maintaining, and distributing the global situation map, providing global information for the execution of intelligent agents.

[0033] In one embodiment, the present invention provides a multi-agent task scheduling method based on a distributed architecture, such as... Figure 2 As shown, it includes the following steps: Step S11: Divide the original task into a multi-dimensional adaptive granularity to form a set of subtasks; Step S12: Evaluate the capabilities of the agents and schedule each subtask in the subtask set to be executed by multiple agents based on the capabilities of the agents; Step S13: Aggregate the subtask execution results of each agent to obtain the final task execution result.

[0034] In one embodiment, step S11, which involves performing multi-dimensional adaptive granular division of the original task, includes vertical decomposition based on business logic, horizontal slicing based on computational complexity, and spatiotemporal alignment based on execution duration.

[0035] In one embodiment, step S12, which evaluates the agent's capabilities, includes: using a resource matching degree function to evaluate the agent's capabilities based on the resources required for the task and the current resource state vector.

[0036] In one embodiment, step S12, which involves scheduling each subtask in the subtask set to multiple agents for execution, includes time-triggered scheduling and event scheduling. The time-triggered scheduling uses a comprehensive scoring mechanism to allocate tasks based on the time of task scheduling, quantifying and integrating three key indicators: resource matching degree, distance affinity, and load idle degree, to achieve optimal agent selection. The event scheduling is driven by specific events triggered by agents to start, stop, execute, or change the state of tasks.

[0037] In one embodiment, the above method further includes: constructing, maintaining and distributing a global situation map using a situation center, providing global information to each agent through the global situation map, so that each agent can maintain real-time status of resources and environment during the execution of sub-tasks.

[0038] In one embodiment, the above method further includes: deploying a master scheduling center and slave centers using a distributed architecture, and using the master scheduling center and slave centers to schedule each subtask to multiple intelligent agents for execution; the master scheduling center is a top-level decision-making and macro-allocation node, and the slave centers are local scheduling and decision-making nodes; the slave centers receive task packages issued by the master scheduling center and allocate subtasks to intelligent agents.

[0039] In one embodiment, after all subtasks are completed in step S13, the main scheduling center or a designated intelligent agent aggregates the execution results of each subtask and triggers a feedback mechanism to optimize the next round of task scheduling.

[0040] In one embodiment, a multi-agent task scheduling method based on a distributed architecture is provided for game-theoretic scenarios. This method achieves task coordination based on a joint counter-strategy, enables dynamic unit deployment, and schedules dynamic task execution based on situational awareness (location, terrain, weather) and events. The steps of this method are as follows: Step 1: Task awareness and registration.

[0041] 1.1) Task Input and Parsing. Based on the system's configured joint countermeasure plan, the data parser parses the task data into standard structured objects and submits them to the scheduling center for registration. The scheduling center collects situational data in real time and stores it in a cache. Task polling is performed every second to determine if the task meets the execution conditions. The scheduling center uses a distributed architecture with a master-slave design, including a master scheduling center and slave centers. Different tasks are distributed to different intelligent agents (workers) based on load balancing.

[0042] Example of a structured object for task input: { "task_id": "T-001", "task_type": "xx task", "priority": "HIGH", "duration": 2, / / Unit: seconds "location": "XX area", "resources": ["U1", "F1"], "dependencies": ["Cannot be executed before T2 completes"], "risk_level": "HIGH", "impact_score": 8.5 } 1.2) Multi-dimensional adaptive granularity division.

[0043] (1) Based on the vertical decomposition of business logic, the business process of complex tasks is decomposed into independent stages with single responsibilities to achieve parallel processing. For example, according to key nodes, the original task is divided into multiple serial or parallel sub-task chains, which are divided into sub-tasks such as "reconnaissance", "planning" and "strike".

[0044] (2) Horizontal slicing based on computational complexity prevents a single subtask from being too large, causing the agent to be blocked for a long time, or too small, causing excessive scheduling overhead. For example, if the amount of reconnaissance data is too large, it can be split into "Region A Reconnaissance" and "Region B Reconnaissance".

[0045] (3) Based on the spatiotemporal alignment of execution time, ensure that the execution cycle of subtasks matches the frequency of environmental changes, and avoid drastic changes in the environment when the task is completed. For example, if “area A reconnaissance” takes too long, it can be split into “frame A-1”, “frame A-2”, etc.

[0046] The final output is a standardized set of atomic subtasks that can be executed in parallel. Through this multi-dimensional and adaptive granularity partitioning, the present invention ensures that tasks are neither too coarse to be executed in parallel on heterogeneous nodes, nor too fine to incur huge communication and scheduling overhead, thus achieving truly efficient, dynamic, and adaptive scheduling.

[0047] Step 2: Assessment and modeling of agent capabilities.

[0048] The intelligent agent reports CPU, memory, and bandwidth usage at regular intervals (e.g., every 5 seconds), creating a capability profile. To achieve intelligent decision-making, tasks and environmental variables are quantified, and an evaluation model, namely the resource matching function, is established. in: : Current resource state vector. The resource state vector refers to the set of quantified values ​​of various available resources held by the agent at time T (time stamp); T represents the timestamp of data collection to ensure the freshness of the state; R represents the set of resource dimensions (such as computing power, storage, communication bandwidth, etc.).

[0049] : Resources required for the task (e.g., [U1, F1]), where R needed This represents the minimum set of resource thresholds necessary to execute a subtask. U1 represents general computing resource requirements (such as the number of CPU cores or memory size required), and F1 represents special function resource requirements (such as specific GPU acceleration or high-bandwidth channels).

[0050] For example: Task T1 requires [U1, F1]. Current status: U1 is available, F1 is available → Resource matching degree = 1. If F1 is unavailable → Resource matching degree = 0.5 (because only part of the requirement is met).

[0051] Step 3: Task allocation and scheduling.

[0052] Task scheduling is divided into two types: time-triggered scheduling and event-triggered scheduling.

[0053] 3.1) Time-triggered scheduling. Time-triggered scheduling combines the timing of task scheduling with a "comprehensive scoring mechanism" for task allocation, quantifying and integrating three key indicators: resource matching degree, distance, and load, to achieve optimal agent selection.

[0054] Comprehensive scoring algorithm: in: Intelligent agent Total priority score (the higher the score, the higher the priority).

[0055] , , Weighting coefficients (e.g.) =0.5, =0.3, =0.2), which can be dynamically adjusted according to the task type (e.g., focusing on distance for urgent tasks and focusing on matching degree for intensive tasks).

[0056] Resource matching degree (base score, 0~1).

[0057] Distance affinity (normalized score, 0-1). Distance affinity is the inverse mapping of the network communication overhead or physical topology distance between the task initiator and the candidate agent. This metric aims to minimize data transmission latency and improve response speed. The closer the distance (the lower the latency), the higher the score. It can be calculated using inverse normalization, as follows: in This represents the distance between the task initiator and the current candidate agent (the i-th candidate agent). The distance between the task initiator and the k-th candidate agent is represented by n, where n is the number of agents.

[0058] Calculate the percentage of the inverse delay of all candidate agents, ensuring that the score is between 0 and 1 and the sum is 1.

[0059] Load idleness (normalized score, 0~1). Load idleness refers to the proportion of resources currently unused by the agent, reflecting its potential to receive new tasks and its resilience. The lower the load (the more idle resources), the higher the score. It can be calculated using the bottleneck resource idle rate method, as follows: This indicates the CPU utilization rate of the intelligent agent. This indicates the memory usage rate of the intelligent agent. This represents the bandwidth utilization rate of the intelligent agent.

[0060] This method takes the resource with the lowest idle rate as the score, which can prevent tasks from being assigned to agents that are about to run out of a resource (such as memory), thus enhancing the stability of the system.

[0061] 3.2) Event Scheduling. Event scheduling refers to the starting, stopping, execution, or state change of tasks not being triggered by a preset time point, but rather driven by a specific event triggered by one or more agents in the system. The core objective is to achieve dynamic, responsive, and context-adaptive task scheduling, making task execution strongly coupled with environmental changes and state evolution, thereby improving system agility and robustness.

[0062] Events in event scheduling mainly include status events, resource events, and status change events. The execution flowchart for event scheduling is as follows: Figure 3 As shown, it includes the following steps: (1) Event triggering: When an agent detects that "the target object has been found" or "an obstacle has appeared in the path", it immediately generates an event object.

[0063] (2) Event broadcast: The event object is broadcast to all agents in the system via the RabbitMQ message queue.

[0064] (3) Event listening: Each agent runs an "event listener" to continuously listen for specified event types.

[0065] (4) Event matching: After an event is received, the system determines whether the event is related to a certain task according to the preset event-task mapping rules, and then matches the event to the target task.

[0066] (5) Task start / stop decision: If the event indicates "target reached", then "task start" is triggered; if the event indicates "resource shortage", then "task pause" is triggered; if the event indicates "task completed", then "task stop" is triggered.

[0067] (6) State update and execution: The task state is updated within the agent and the executor responds immediately, realizing the "perception-response-execution" closed loop.

[0068] The essence of event scheduling is to achieve task start / stop decisions through "relationship modeling" and "state coordination" between agents. After an event is triggered, the system determines, through matching rules, "which agent" should be the "decision-making body for task start / stop" or the "executing body". To achieve agent matching, the system maintains an event-task-agent matching rule base.

[0069] Step 4: Task execution and status synchronization.

[0070] Once the task meets the execution conditions, the task execution is triggered, and the locally cached agent is scheduled. After the task is completed, the scheduling center will use the RabbitMQ message queue to broadcast the task execution status to the system server. After parsing, the system server will return the data to the client to display the task execution status and task execution process log.

[0071] Step 5: Result aggregation and task closure.

[0072] During task execution, the scheduling center aggregates resource and target data in real time and feeds back standardized structured data to each client, forming a unified "system status snapshot" or "decision basis".

[0073] After implementation, this technical solution reduced task completion time by 35%, increased system availability to 90%, and eliminated single points of failure. Key aspects involved in implementation included: distributed architecture, event triggering, and a situational awareness center.

[0074] Key Point 1: The beneficial effects of distributed architecture are shown in Table 1.

[0075] Table 1 Key Point 2: The beneficial effects of event triggering are shown in Table 2.

[0076] Table 2 Key Point 3: A beneficial effect of the situation center is the achievement of consistency in situation handling. The value and role of consistent situation handling include: (1) Resolving resource conflicts. When all agents have a consistent understanding of resource distribution and usage status, "duplicate requests" or "resource waste" can be avoided.

[0077] (2) Improve task scheduling efficiency. Based on a unified situation, the scheduling center can accurately determine task feasibility, resource requirements and execution paths, thus avoiding ineffective scheduling.

[0078] (3) Enhance system robustness. When some agents fail or lose information, the remaining agents can still reason and make decisions based on the "consistent situation".

[0079] (4) Support cross-domain collaboration. In multi-domain (such as drones, robots, edge computing) collaborative tasks, unified situational awareness is the foundation for achieving "collaborative perception and collaborative decision-making".

[0080] Another embodiment of the present invention provides a multi-agent task scheduling system based on a distributed architecture, such as... Figure 4 As shown, it includes: The task partitioning module 20 is used to partition the original task into a multi-dimensional adaptive granularity to form a set of subtasks; Task scheduling module 21 is used to evaluate the capabilities of the agents and schedule each subtask in the subtask set to be executed by multiple agents based on the capabilities of the agents. The aggregation module 22 is used to aggregate the subtask execution results of each agent to obtain the final task execution result.

[0081] The above division of modules is merely illustrative. In practical applications, the functions described above can be assigned to different functional modules as needed to complete all or part of the functions described in the aforementioned method. The specific working process of each module can be referred to the corresponding process in the aforementioned method embodiments, and will not be repeated here. Each of the above modules can be implemented entirely or partially through software, hardware, or a combination thereof.

[0082] Another embodiment of the present invention provides a computer device (computer, server, smartphone, etc.) including a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing steps of the method of the present invention.

[0083] Another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk) that stores a computer program, which, when executed by a computer, implements the steps of the method of the present invention.

[0084] Another embodiment of the present invention provides a computer program product, the computer program product including a computer program, which, when executed by a computer, implements the steps of the method of the present invention.

[0085] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and to implement it accordingly. Those skilled in the art will understand that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the content disclosed in the embodiments of this specification; the scope of protection of the present invention is defined by the claims.

Claims

1. A multi-agent task scheduling method based on a distributed architecture, characterized in that, Includes the following steps: The original task is divided into subtasks by multi-dimensional adaptive granularity. Evaluate the capabilities of the agents and, based on those capabilities, schedule each subtask in the subtask set to be executed by multiple agents. The results of the subtasks performed by each agent are aggregated to obtain the final task execution result.

2. The method according to claim 1, characterized in that, The process of dividing the original task into multi-dimensional adaptive granularities includes vertical decomposition based on business logic, horizontal slicing based on computational complexity, and spatiotemporal alignment based on execution duration.

3. The method according to claim 1, characterized in that, The evaluation of the agent's capabilities includes: using a resource matching degree function to evaluate the agent's capabilities based on the resources required for the task and the current resource state vector.

4. The method according to claim 1, characterized in that, The process of scheduling each subtask in the subtask set to multiple agents for execution includes time-triggered scheduling and event-triggered scheduling. Time-triggered scheduling uses a comprehensive scoring mechanism to allocate tasks based on the timing of task scheduling, quantifying and integrating three key indicators: resource matching degree, distance affinity, and load idle degree, to achieve optimal agent selection. Event-triggered scheduling drives the start, stop, execution, or state change of tasks through specific events triggered by agents.

5. The method according to claim 1, characterized in that, The global situation map is constructed, maintained, and distributed using the situation center. The global situation map provides global information to each agent, enabling each agent to maintain real-time awareness of resources and environment during the execution of sub-tasks.

6. The method according to claim 1, characterized in that, A distributed architecture is adopted to deploy a master scheduling center and slave centers. The master scheduling center and slave centers are used to schedule each subtask to multiple intelligent agents for execution. The master scheduling center is the top-level decision-making and macro-allocation node, and the slave centers are local scheduling and decision-making nodes. The slave centers receive task packages issued by the master scheduling center and allocate subtasks to intelligent agents.

7. The method according to claim 6, characterized in that, After all subtasks are completed, the main scheduling center or a designated agent aggregates the execution results of each subtask and triggers a feedback mechanism to optimize the next round of task scheduling.

8. A multi-agent task scheduling system based on a distributed architecture, characterized in that, include: The task partitioning module is used to divide the original task into subtask sets by multi-dimensional adaptive granularity. The task scheduling module is used to evaluate the capabilities of the agents and schedule each subtask in the subtask set to be executed by multiple agents based on the capabilities of the agents. The aggregation module is used to aggregate the subtask execution results of each agent to obtain the final task execution result.

9. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a computer, implements the method according to any one of claims 1 to 7.

Images