CN-122022418-A - Task allocation method, medium and equipment for layered multi-agent

CN122022418ACN 122022418 ACN122022418 ACN 122022418ACN-122022418-A

Abstract

The invention relates to the technical field of task distribution, in particular to a task distribution method, medium and equipment of layered multi-agent, which are used for carrying out subtask decomposition, dependency analysis and execution structure judgment through task state characteristics, comprising a first task graph of a serial execution sequence, improving the task distribution accuracy, distributing optimal target agent to each subtask and constructing an efficient serial agent execution chain through matching subtask demand characteristics and agent capability characteristics, improving the basic efficiency and quality of subtask execution, and timely adjusting the task graph and agent configuration by collecting agent execution state data to sense the subtask execution abnormality in real time and timely adjust the task graph and agent configuration, thereby avoiding the overall task failure caused by the local execution problem, guaranteeing the stability and continuity of complex serial task execution and improving the completion quality and execution efficiency of the complex task.

Inventors

ZHU JIE
ZHANG SEN
Bai Qunwei

Assignees

中科雨辰科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (10)

1. A method for task allocation of a layered multi-agent, the method comprising the steps of: s1, analyzing task state characteristics corresponding to a target task to obtain a plurality of subtasks, logic dependency relations among the subtasks and an execution structure, wherein the execution structure is in a series mode, a parallel mode or a mixed mode; s2, if the execution structure is in a serial mode, sequencing all the subtasks according to the logic dependency relationship, and generating a first task graph containing serial execution sequences, wherein the serial execution sequences indicate the processed sequence of the subtasks; S3, screening to obtain target agents corresponding to each subtask according to the task demand characteristics of each subtask and the capability characteristics of each initial agent in the first task graph, and constructing to obtain a serial agent execution chain; S4, controlling the serial intelligent agent execution chain to execute each subtask of the first task graph according to the serial execution sequence to obtain execution state data and task execution results corresponding to each target intelligent agent; S5, judging whether to trigger a re-planning operation according to the execution state data corresponding to all the target agents; and S6, if the re-planning operation is not triggered, obtaining a target execution result according to task execution results corresponding to all target agents, otherwise, updating the first task graph according to execution state data up to the trigger time, the completed task execution result and the residual unexecuted subtasks, and returning to the execution step S3 by using the updated first task graph.
2. The task allocation method of a layered multi-agent according to claim 1, wherein S1 comprises the steps of: S11, performing natural language processing on the target task, and extracting semantic information of the target task, wherein the semantic information comprises task intention, key entities and constraint conditions; S12, according to the semantic information, retrieving historical execution records and context information related to the target task; S13, fusion encoding is carried out on the semantic information, the history execution record and the context information, so that task state characteristics of the target task are obtained; S14, inputting the task state characteristics into a task decomposition network in a meta-decision maker, and outputting a plurality of subtasks and semantic descriptions of each subtask; S15, predicting to obtain a logic dependency relationship between any two subtasks according to semantic descriptions of the subtasks and a relationship prediction network in the meta-decision maker, wherein the logic dependency relationship comprises sequential dependency, parallel feasibility or no dependency; S16, judging the execution structure of the target task through a structure decision maker in the meta-decision maker according to the logic dependency relationship among the subtasks.
3. The task allocation method of a layered multi-agent according to claim 2, wherein S16 comprises the steps of: s161, constructing a logic dependency graph which takes a subtask as a node and takes a dependency relationship as a directed edge according to the logic dependency relationship between any two subtasks; S162, carrying out structural analysis on the logic dependency graph, and extracting to obtain key topological features, wherein the key topological features comprise node output degree, node input degree, graph connectivity, key path length, graph depth and parallel clusters; s163, performing mode classification according to the key topological features and a preset judgment rule in the structure decision maker, and outputting an execution structure of the target task.
4. The task allocation method of layered multi-agent according to claim 1, wherein S3 comprises the steps of: S31, extracting task demand characteristics of each subtask and capability characteristics of each initial agent, wherein the task demand characteristics comprise function demand characteristics, precision demand characteristics and calculation complexity demand characteristics, and the capability characteristics comprise function label characteristics, historical execution accuracy and resource occupancy rate; S32, inputting task demand characteristics of a current subtask and capability characteristics of any initial agent into a preset weight distribution network aiming at any subtask in the first task graph, and acquiring an adaptive weight between the current subtask and any initial agent, wherein the preset weight distribution network carries out training optimization based on execution state data of the agents in a historical task; s33, determining the initial agent corresponding to the maximum adaptive weight as the target agent corresponding to the current subtask; s34, traversing all the subtasks in the first task graph to obtain a target agent corresponding to each subtask; And S35, sequentially connecting target agents corresponding to all the subtasks according to the serial execution sequence to obtain the serial agent execution chain.
5. The method for task allocation of layered multi-agent according to claim 1, wherein S4 comprises the steps of: s41, controlling a first target agent in the serial agent execution chain to execute the corresponding allocated subtasks to obtain a corresponding task execution result; S42, collecting execution state data of each target intelligent agent in real time, wherein the execution state data comprise an output quality index, an execution efficiency index and a resource consumption index; S43, according to the serial execution sequence, after the execution of the preceding subtasks is finished, transmitting the corresponding task execution result to a target agent corresponding to the following subtasks according to a preset communication protocol; S44, controlling target agents of the follow-up subtasks to execute the corresponding subtasks according to the received task execution results, and repeatedly executing the steps S42-S43 until all the subtasks are executed, so as to obtain execution state data and task execution results corresponding to the target agents.
6. The method for assigning tasks to layered multi-agent according to claim 5, wherein S5 comprises the steps of: And monitoring the execution state data of each target intelligent agent, and triggering a reprofiling operation if the output quality index of any target intelligent agent is smaller than a preset quality threshold, the execution efficiency index is smaller than a preset efficiency threshold or the resource consumption index is larger than a preset consumption threshold.
7. The method for assigning tasks of layered multi-agent according to claim 6, wherein S6 comprises the steps of: s61, if the re-planning operation is triggered, the execution state data, the completed task execution result and the residual non-executed subtasks at the trigger moment are input to a preset re-planning decision network after being encoded, and a first task adjustment scheme is obtained, wherein the first task adjustment scheme comprises the steps of skipping the subtasks, replacing target agents, adjusting the logic dependency relationship among the residual subtasks or inserting new error correction subtasks; s62, updating the first task graph according to the first task adjustment scheme, and generating an updated first task graph; And S63, returning to the step S3 based on the updated first task graph until the re-planning operation is not triggered.
8. The method for task allocation of layered multi-agent according to claim 1, wherein S6 comprises the steps of: and integrating task execution results of all corresponding target agents when the re-planning operation is not triggered according to the corresponding serial execution sequence to form the target execution result.
9. A non-transitory computer readable storage medium having at least one instruction or at least one program stored therein, wherein the at least one instruction or the at least one program is loaded and executed by a processor to implement the method of task allocation of a layered multi-agent according to any one of claims 1-8.
10. An electronic device comprising a processor and the non-transitory computer readable storage medium of claim 9.

Description

Task allocation method, medium and equipment for layered multi-agent Technical Field The present invention relates to the field of task allocation technologies, and in particular, to a task allocation method, medium, and device for a layered multi-agent. Background With the rapid development of artificial intelligence and multi-intelligent system technology, the intelligent allocation and execution of complex tasks becomes an important research point in the field. In the scenes of industrial process control, intelligent question-answering, cross-domain information integration and the like, complex tasks often have the characteristics of multiple steps and strong dependence, and the complex tasks are required to be decomposed into a plurality of subtasks and executed according to a specific sequence. The existing serial distribution method is generally driven by adopting a static rule, namely, the disassembly and the execution sequence setting of complex tasks are completed through manually presetting a fixed flow template, the tasks are not subjected to deep structural analysis, and the internal logic dependency relationship among the subtasks cannot be mined, so that the sequential execution sequence of the subtasks lacks uniform structural representation, and the subtask execution sequence is easy to be inconsistent with the actual business logic. Meanwhile, the existing serial distribution method does not establish a quantitative matching system of task demands and the capabilities of the intelligent agents, and does not combine with a state data optimization matching strategy in the execution process of the intelligent agents, when the capabilities of the intelligent agents change due to resource consumption and performance fluctuation, the intelligent agents can only continue to advance according to a fixed flow, the adaptive intelligent agents cannot be replaced in time, and the overall task execution quality and efficiency are greatly reduced. In addition, the existing serial allocation method does not construct a dynamic re-planning mechanism, and when aiming at emergencies such as abnormal execution of an agent, resource conflict and the like, the configuration of a subtask sequence and the agent is difficult to quickly adjust, only tasks can be interrupted or low-quality results can be accepted, and the requirements of high precision and high efficiency under complex scenes are difficult to meet. Therefore, how to construct a layered multi-agent serial task allocation method capable of sensing the execution state of agents and supporting dynamic re-planning becomes a problem to be solved. Disclosure of Invention Aiming at the technical problems, the technical scheme adopted by the invention is a task allocation method of a layered multi-agent, which comprises the following steps: s1, analyzing task state characteristics corresponding to a target task to obtain a plurality of subtasks, logic dependency relations among the subtasks and an execution structure, wherein the execution structure is in a series mode, a parallel mode or a mixed mode. S2, if the execution structure is in a serial mode, all the subtasks are ordered according to the logic dependency relationship, and a first task graph containing serial execution sequences is generated, wherein the serial execution sequences indicate the processed sequence of the subtasks. S3, screening to obtain target agents corresponding to each subtask according to the task demand characteristics of each subtask and the capability characteristics of each initial agent in the first task graph, and constructing to obtain a serial agent execution chain. S4, controlling the serial intelligent agent execution chain to execute each subtask of the first task graph according to the serial execution sequence, and obtaining execution state data and task execution results corresponding to each target intelligent agent. And S5, judging whether to trigger the re-programming operation according to the execution state data corresponding to all the target agents. And S6, if the re-planning operation is not triggered, obtaining a target execution result according to task execution results corresponding to all target agents, otherwise, updating a first task graph according to execution state data up to the trigger time, the completed task execution result and the rest of non-executed subtasks, and returning to the execution step S3 by the updated first task graph. The invention also provides a non-transitory computer readable storage medium, wherein at least one instruction or at least one section of program is stored in the non-transitory computer readable storage medium, and the at least one instruction or the at least one section of program is loaded and executed by a processor to realize the task allocation method of the layered multi-agent. The invention also provides an electronic device comprising a processor and the non-transitory computer readable storage medium d