Search

CN-121985040-A - Centralized multi-agent cooperative scheduling method, system, terminal and storage medium for universal intelligent networking scene

CN121985040ACN 121985040 ACN121985040 ACN 121985040ACN-121985040-A

Abstract

The invention discloses a non-centralized multi-agent cooperative scheduling method, a system, a terminal and a storage medium for a universal intelligent joint scene, wherein the method comprises the steps of receiving original task data from universal intelligent joint equipment, and decomposing the original task data into a plurality of atomic subtasks based on a dynamic task decomposition algorithm; the resource scheduling decision-making agent generates resource scheduling decisions for the atomic subtasks according to the real-time state of the system, distributes the atomic subtasks carrying the scheduling decisions to corresponding execution agents and computing nodes through a non-centralized P2P message bus, cooperatively executes the atomic subtasks based on the resource scheduling decisions by the execution agents and the computing nodes, and synthesizes and outputs execution results. The invention breaks through the performance bottleneck of the centralized scheduling architecture, and realizes the extremely low-delay cooperative scheduling of massive concurrent tasks and the real-time guarantee of seamless cooperation of cross-protocol equipment and high-priority tasks.

Inventors

  • LU WEICHAO

Assignees

  • 深圳开鸿数字产业发展有限公司

Dates

Publication Date
20260505
Application Date
20251225

Claims (20)

  1. 1. The centerless multi-agent cooperative scheduling method for the intelligent networking scene is characterized by comprising the following steps of: Receiving original task data from a universal intelligent joint device, and decomposing the original task data into a plurality of atomic subtasks based on a dynamic task decomposition algorithm; generating resource scheduling decisions for the atomic subtasks by a resource scheduling decision agent according to the real-time state of the system; Distributing each atomic subtask carrying a scheduling decision to a corresponding execution agent and a computing node through a non-centralized P2P message bus; And the execution agent and the computing node cooperatively execute each atomic subtask based on the resource scheduling decision, and synthesize and output an execution result.
  2. 2. The non-centralized multi-agent cooperative scheduling method for the intelligent networking scene of claim 1, the method is characterized in that the receiving of the original task data from the universal intelligent joint equipment specifically comprises the following steps: receiving original task data directly uploaded by the universal intelligent connection device through a P2P message bus of a Matrix centerless architecture; The universal intelligent device is directly connected to the P2P message bus in a point-to-point mode and at least comprises one of an industrial sensor, a transaction processing terminal and a city camera; and performing preliminary feature extraction and encapsulation on the received original task data by using the equipment state sensing Agent to form a standardized task data object.
  3. 3. The method for non-centralized multi-agent coordination scheduling for a universal intelligent joint scene according to claim 1, wherein the decomposing the original task data into a plurality of atomic subtasks based on a dynamic task decomposition algorithm specifically comprises: Defining the original task data as a target task to be decomposed, and acquiring semantic entropy of the target task and a preset complexity coefficient; Calculating the number of atomic subtasks through a preset function according to the semantic entropy and the complexity coefficient, wherein the preset function is used for carrying out multi-granularity analysis and load balancing modeling according to task complexity and system load; based on the decomposition strategy parameters formed by the semantic entropy and the complexity coefficient, performing multi-granularity analysis and dependency modeling on the target task to obtain an analysis modeling result; And dynamically disassembling the target task into a plurality of atomic subtasks which can be independently or parallelly executed according to the number according to the analysis modeling result.
  4. 4. The non-centralized multi-agent cooperative scheduling method for a universal intelligent networking scenario of claim 1, wherein the resource scheduling decision agent is a component of a non-centralized agent cooperative network, the non-centralized agent cooperative network further comprising: The equipment state sensing intelligent body is used for dynamically collecting and reporting equipment data characteristics; And the service dynamic arrangement agent is used for binding the service instance corresponding to the atomic subtask with the underlying physical resource or the device driving container according to the resource scheduling decision.
  5. 5. The method for non-centralized multi-agent cooperative scheduling for a universal intelligent joint scenario according to claim 4, wherein the resource scheduling decision-making agent generates a resource scheduling decision for each of the atomic subtasks according to a real-time state of a system, specifically comprising: Receiving, by the resource scheduling decision agent, a real-time load indicator reported by the device state awareness agent through the P2P message bus, wherein the indicator at least includes one of device data throughput, GPU utilization, container start delay, and network round trip delay; The resource scheduling decision-making agent calculates and generates a scheduling decision comprising a target execution node, a priority and a resource quota constraint for each atomic subtask through a row-level scheduling algorithm according to the received real-time load index and the attribute of each atomic subtask; And the resource scheduling decision agent encapsulates the generated scheduling decision into a scheduling instruction object capable of serialization, and distributes the scheduling instruction object to the corresponding service dynamic scheduling agent and the execution node through the P2P message bus.
  6. 6. The non-centralized multi-agent cooperative scheduling method for the universal intelligent joint scene according to claim 5, wherein the decision logic of the row-level scheduling algorithm is realized based on a resource scheduling decision matrix, and decisions corresponding to different types of atomic subtasks are predefined in the resource scheduling decision matrix; The method comprises the steps of adopting an event-driven scheduling strategy aiming at equipment perception atomic subtasks, and taking equipment data throughput as a key feedback index; aiming at LLM reasoning atomic subtasks, a dynamic batch processing scheduling strategy is adopted, and the GPU utilization rate is used as a key feedback index; aiming at a tool call atomic subtasks, adopting an affinity scheduling strategy, and taking the starting delay of a container as a key feedback index; Aiming at the edge computing atomic subtasks, a geographic position sensing scheduling strategy is adopted, and network round trip delay is used as a key feedback index.
  7. 7. The method for non-centralized multi-agent cooperative scheduling for a universal intelligent joint scenario according to claim 5, wherein after the resource scheduling decision agent generates a resource scheduling decision for each of the atomic subtasks according to a real-time state of the system, the method further comprises: And continuously monitoring the real-time state of the system, when the local state deviation degree is detected to exceed a preset threshold value through a hierarchical control coordination mechanism or a high-priority task preemption event is identified, immediately interrupting the current scheduling decision flow, and regenerating an updated scheduling decision through the row-level scheduling algorithm based on the latest global state.
  8. 8. The non-centralized multi-agent coordination scheduling method for a universal intelligent networking scenario of claim 7, wherein the detection is performed by a hierarchical control coordination mechanism, specifically comprising: Constructing a double-layer scheduling state machine consisting of a global coordinator and a plurality of local executors, wherein the global coordinator is used for continuously converging the atomic subtask execution state vectors reported by the local executors and calculating the state offset of the local executors; when the state offset of any local executor exceeds a preset threshold value or a high-priority task preemption event is identified, the state transfer function of the global coordinator switches the system scheduling state to a rebalancing state and triggers a global instant rescheduling mechanism; The global instant rescheduling mechanism reconstructs, redistributes resources and adjusts an execution path of the affected atomic subtasks based on the latest system state converged by the global coordinator, and generates a new scheduling decision to redistribute execution.
  9. 9. The method for non-centralized multi-agent cooperative scheduling for a universal intelligent networking scenario according to claim 1, wherein the distributing, through a non-centralized P2P message bus, each of the atomic subtasks carrying scheduling decisions to a corresponding executing agent and a computing node specifically comprises: Packaging the atomic subtasks and the scheduling decisions carried by the atomic subtasks into task message objects which are in a unified format and can be sequenced, and issuing the task message objects to the P2P message bus; And transmitting the task message object to a corresponding execution agent and a corresponding computing node by the P2P message bus according to the target execution node identifier contained in the scheduling decision in a point-to-point or multipoint-to-multipoint mode and a mixed routing strategy based on the content and the node identifier.
  10. 10. The method of claim 9, wherein the task message object further includes a service instance identifier or a device driver container identifier bound by a service dynamic orchestration agent for loading and execution by the execution agent and computing nodes.
  11. 11. The method for non-centralized multi-agent cooperative scheduling for a universal intelligent joint scenario according to claim 1, wherein the executing agent and the computing node cooperatively execute each atomic subtask based on the resource scheduling decision, and synthesize and output an execution result, specifically comprising: Loading a corresponding service instance or equipment driving container by the execution agent and the computing node according to the target node and the resource constraint specified in the resource scheduling decision, and executing a corresponding atomic subtask; Each executing node exchanges executing state and intermediate data in real time through the P2P message bus so as to realize cooperative execution; and after all the atomic subtasks are executed, performing time-sensitive weighted fusion on the results of all the atomic subtasks through an asynchronous aggregation function to obtain the synthesized final output.
  12. 12. The method for non-centralized multi-agent coordination scheduling for a universal intelligent joint scenario according to claim 11, wherein after the executing nodes exchange the executing state and the intermediate data in real time through the P2P message bus to realize the coordination execution, the method further comprises: if any executing node reports that the execution fails or overtime, triggering a loop verification mechanism to judge and locate errors, and obtaining a verification result; and based on the verification result, reconstructing and redistributing the failed atomic subtasks through an instant rescheduling mechanism.
  13. 13. The method for non-centralized multi-agent coordination scheduling for a universal intelligent joint scene according to claim 12, wherein the step of performing time-sensitive weighted fusion on the results of each atomic subtask through an asynchronous aggregation function to obtain a synthesized final output further comprises: Inputting the synthesized final output into the circulation verification mechanism for multidimensional verification to obtain a verification result; And if the verification is passed, the synthesized final output and the related monitoring index are subjected to persistent storage and are issued to the outside through a standardized data interface.
  14. 14. The centralized multi-agent cooperative scheduling system for the intelligent networking scene is characterized by comprising the following components: The data acquisition and decomposition module is used for receiving the original task data from the universal intelligent joint equipment and decomposing the original task data into a plurality of atomic subtasks based on a dynamic task decomposition algorithm; the resource scheduling decision generation module is used for generating resource scheduling decisions for the atomic subtasks by a resource scheduling decision agent according to the real-time state of the system; the subtask distribution module is used for distributing each atomic subtask carrying a scheduling decision to a corresponding execution agent and a corresponding computing node through a non-centralized P2P message bus; And the subtask execution module is used for cooperatively executing each atomic subtask by the execution agent and the computing node based on the resource scheduling decision, and synthesizing and outputting an execution result.
  15. 15. The non-centralized multi-agent collaborative scheduling system for a case of a wisdom alliance scenario according to claim 14, wherein the data acquisition and decomposition module includes a data acquisition encapsulation unit, a task number determination unit, and a subtask splitting unit; the acquisition and encapsulation unit is used for receiving original task data directly uploaded by the universal intelligent device through a P2P message bus of a Matrix centerless architecture, and carrying out preliminary feature extraction and encapsulation on the received original task data by the device state sensing Agent to form a standardized task data object; the task number determining unit is used for defining the original task data as a target task to be decomposed, acquiring semantic entropy of the target task and a preset complexity coefficient, and calculating the number of atomic subtasks through a preset function according to the semantic entropy and the complexity coefficient; The subtask splitting unit is used for carrying out multi-granularity analysis and dependency modeling on the target task based on a decomposition strategy parameter formed by the semantic entropy and the complexity coefficient to obtain an analysis modeling result, and dynamically splitting the target task into a plurality of atomic subtasks which can be independently or parallelly executed according to the number according to the analysis modeling result.
  16. 16. The non-centralized multi-agent collaborative scheduling system for a case of a universal intelligent joint according to claim 14, wherein the resource scheduling decision-making module comprises a load index receiving unit, a scheduling decision-making unit and a packaging executing unit; The load index receiving unit is used for receiving the real-time load index reported by the equipment state sensing agent by the resource scheduling decision agent through the P2P message bus, and the index at least comprises one of equipment data throughput, GPU utilization rate, container starting delay and network round trip delay; the scheduling decision generation unit is used for calculating and generating a scheduling decision comprising a target execution node, a priority and a resource quota constraint for each atomic subtask through a row-level scheduling algorithm according to the received real-time load index and the attribute of each atomic subtask by the resource scheduling decision agent; The packaging execution unit is used for packaging the generated scheduling decision into a scheduling instruction object capable of serialization by the resource scheduling decision agent, and distributing the scheduling decision to the corresponding service dynamic arrangement agent and execution nodes through the P2P message bus.
  17. 17. The non-centralized multi-agent collaborative scheduling system for a case of a wisdom alliance scenario according to claim 14, wherein the subtask distribution module includes a decision encapsulation sending unit and a transmission execution unit; The decision encapsulation sending unit is used for encapsulating the atomic subtasks and the scheduling decisions carried by the atomic subtasks into task message objects which are in a unified format and can be serialized, and publishing the task message objects to the P2P message bus; The transmission execution unit is used for transmitting the task message object to the corresponding execution agent and the corresponding computing node by the P2P message bus in a point-to-point or multipoint-to-multipoint mode according to the target execution node identifier contained in the scheduling decision and based on the mixed routing strategy of the content and the node identifier.
  18. 18. The multi-agent co-scheduling system without centralization for a case of a wisdom alliance scenario of claim 14, wherein the subtask execution module comprises a corresponding execution unit, a co-execution unit, and a synthetic output unit; The corresponding execution unit is used for loading a corresponding service instance or equipment driving container by the execution agent and the computing node according to the target node and the resource constraint designated in the resource scheduling decision, and executing a corresponding atomic subtask; the cooperative execution unit is used for each execution node to exchange the execution state and the intermediate data in real time through the P2P message bus so as to realize cooperative execution; And the synthesis output unit is used for carrying out time-sensitive weighted fusion on the results of all the atomic subtasks through an asynchronous aggregation function after all the atomic subtasks are executed, so as to obtain synthesized final output.
  19. 19. A terminal comprising a memory, a processor and a non-centralized multi-agent co-scheduler for a wisdom-alliance scenario stored on the memory and operable on the processor, the non-centralized multi-agent co-scheduler for a wisdom-alliance scenario implementing the steps of the non-centralized multi-agent co-scheduler for a wisdom-alliance scenario of any of claims 1-13 when executed by the processor.
  20. 20. A computer readable storage medium, wherein the computer readable storage medium stores a non-centralized multi-agent collaborative scheduler for a wisdom-oriented scenario, which when executed by a processor, implements the steps of the non-centralized multi-agent collaborative scheduling method for a wisdom-oriented scenario of any of claims 1-13.

Description

Centralized multi-agent cooperative scheduling method, system, terminal and storage medium for universal intelligent networking scene Technical Field The invention relates to the technical field of multi-agent coordination scheduling, in particular to a non-centralized multi-agent coordination scheduling method, system, terminal and computer readable storage medium for a universal intelligent networking scene. Background With the rapid development of the universal intelligent technology, the number of devices and data types accessed in scenes such as industrial Internet of things, smart cities and the like are rapidly increased, and the task execution environment shows the characteristics of high concurrency, strong real-time, cross-domain isomerism and the like. To address the collaborative computing needs in such complex scenarios, multi-agent systems (Multi-AGENT SYSTEM, MAS) are widely studied and applied as a distributed intelligent collaborative paradigm. The system completes tasks cooperatively through a plurality of agents with autonomous decision and interaction capability, and has the advantages of flexibility and robustness in a scene with limited resources or dynamic change. However, most current multi-agent collaborative systems still implement task scheduling and resource coordination based on a centralized or hierarchical centralized architecture. The architecture relies on a single or a few central nodes to carry out global state maintenance and decision making, and although the architecture is feasible in a small-and-medium-scale scene, the architecture exposes significant limitations in a high concurrency and high dynamic scene of the universal intelligent link. Although research attempts are made to improve the system through load balancing, dynamic priority queues and other modes, the problems of inherent expansibility, high response delay, difficult cross-domain collaboration and the like of a centralized architecture are not fundamentally solved, and the actual deployment and application effects of a large-scale intelligent agent collaboration system in a universal intelligent networking scene are restricted. Accordingly, the prior art is still in need of improvement and development. Disclosure of Invention The invention mainly aims to provide a non-centralized multi-agent cooperative scheduling method, a system, a terminal and a computer readable storage medium for a universal intelligent networking scene, and aims to solve the problems of resource scheduling bottleneck, low cooperative efficiency and high system response delay when a traditional centralized scheduling architecture processes high concurrent, heterogeneous and real-time tasks in the prior art. In order to achieve the above purpose, the invention provides a centralized-free multi-agent cooperative scheduling method for a universal intelligent scene, which comprises the following steps: Receiving original task data from a universal intelligent joint device, and decomposing the original task data into a plurality of atomic subtasks based on a dynamic task decomposition algorithm; generating resource scheduling decisions for the atomic subtasks by a resource scheduling decision agent according to the real-time state of the system; Distributing each atomic subtask carrying a scheduling decision to a corresponding execution agent and a computing node through a non-centralized P2P message bus; And the execution agent and the computing node cooperatively execute each atomic subtask based on the resource scheduling decision, and synthesize and output an execution result. Optionally, the method for centreless multi-agent coordination scheduling for a case of a smart phone, wherein the receiving the original task data from the case of the smart phone specifically includes: receiving original task data directly uploaded by the universal intelligent connection device through a P2P message bus of a Matrix centerless architecture; The universal intelligent device is directly connected to the P2P message bus in a point-to-point mode and at least comprises one of an industrial sensor, a transaction processing terminal and a city camera; and performing preliminary feature extraction and encapsulation on the received original task data by using the equipment state sensing Agent to form a standardized task data object. Optionally, the method for non-centralized multi-agent coordination scheduling for a universal intelligent scene, wherein the method for decomposing the original task data into a plurality of atomic subtasks based on a dynamic task decomposition algorithm specifically includes: defining the original task data as a target task to be decomposed, and acquiring the target task Semantic entropy and a preset complexity coefficient; Calculating the number of atomic subtasks through a preset function according to the semantic entropy and the complexity coefficient, wherein the preset function is used for carrying out multi-granular