Self-evolution multi-agent cooperation method and system based on dynamic cognitive map

Abstract

This invention discloses a self-evolving multi-agent collaborative method and system based on a dynamic cognitive graph. The method includes constructing a dynamic cognitive graph and modeling the reasoning process as a directed acyclic graph; initializing a multi-agent collaborative network; executing sub-tasks and dynamically reconstructing the cognitive graph using execution feedback; evaluating the execution process using a case-based reasoning evolutionary mechanism, generating structured cases, and encoding and storing the structured cases in a case library for experience reuse in subsequent tasks; the multi-agent collaborative network adopts a hybrid collaborative architecture, communicating between agents through a blackboard system; monitoring metacognition by real-time monitoring of the confidence level of the reasoning process, and using an arbitration mechanism to arbitrate conflicts when multiple agents reach contradictory conclusions; outputting the execution and evaluating the system performance. This invention significantly improves the processing capability of complex long-link reasoning tasks and the system's continuous learning capability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a self-evolving multi-agent collaboration method and system based on dynamic cognitive graphs. Background Technology

[0002] Currently, research on agents centered on Large Language Models (LLMs) is at a critical turning point, evolving from "intuitive fast thinking (System 1)" to "logical slow thinking (System 2)". Although existing multi-agent frameworks (such as AutoGen and MetaGPT) have shown certain collaborative capabilities under standard operating procedures (SOPs), they still face two core bottlenecks when dealing with complex open-domain problems: "cognitive rigidity" and "experience gap".

[0003] Regarding cognitive rigidity, traditional multi-agent collaboration relies heavily on predefined static workflows. Once the task environment undergoes unexpected changes, the system is prone to getting stuck in a dead loop or experiencing hallucinations. While existing Graph of Thoughts introduces graph structures, their graphs are typically statically defined before the task begins, lacking the ability to dynamically reconstruct the graph structure based on environmental feedback at runtime, making it difficult to cope with sudden variables in open-domain environments.

[0004] Regarding experiential gaps, existing intelligent agents generally suffer from a "memoryless" defect, where the reasoning trajectory is reset after each task, preventing the system from learning from past successes and failures. Existing reflection mechanisms (such as Reflexion) are mostly limited to short-term optimization within a task, lacking mechanisms for long-term experience storage and retrieval across tasks. How to transform unstructured natural language reflection into a structured, searchable case library remains a key challenge for achieving lifelong learning for intelligent agents. Summary of the Invention

[0005] The purpose of this invention is to provide a self-evolving multi-agent cooperation method and system based on dynamic cognitive graphs to solve the problems mentioned in the background art.

[0006] The present invention provides the following technical solution: a self-evolving multi-agent cooperation method based on dynamic cognitive graphs, comprising the following operational steps:

[0007] Step S1: Construct a dynamic cognitive graph, modeling the reasoning process as a directed acyclic graph.

[0008] Preferably, the nodes of the directed acyclic graph represent reasoning states or subtasks, and the edges represent state transitions or dependencies. Constructing a dynamic cognitive graph includes: defining the structure of the dynamic cognitive graph; receiving the task description (Task) input by the user and using a planning agent to parse and decompose the task; and dynamically reconstructing the cognitive graph based on execution feedback during task execution.

[0009] Preferably, dynamically reconstructing the cognitive graph based on execution feedback includes: when a subtask is executed successfully, marking the corresponding node as completed and activating the successor node; when a subtask fails, triggering a backtracking mechanism and adding alternative path nodes to the cognitive graph; when a new sub-problem is discovered during execution, dynamically inserting a new node into the cognitive graph and establishing a dependency relationship with existing nodes; and when redundant or invalid paths are detected, pruning and optimizing the cognitive graph by deleting the corresponding nodes and edges.

[0010] Step S2: Initialize a multi-agent cooperative network including a planning agent, an execution agent, and a reflective agent.

[0011] Preferably, the planning agent includes:

[0012] The task parsing module is used to understand the macro-level task objectives input by the user;

[0013] The task decomposition module uses the thinking chain method to break down macro-level tasks into structured sub-problem trees;

[0014] The graph generation module converts the sub-problem tree into an initial cognitive graph in the form of a directed acyclic graph;

[0015] The path planning module plans the execution order based on the priority and dependencies of subtasks.

[0016] Preferably, the executing agent includes:

[0017] The tool invocation module is used to invoke external tools to perform specific operations.

[0018] The information retrieval module is used to obtain the required information from external knowledge sources;

[0019] The status reporting module is used to feed back the execution status and results to the cognitive graph;

[0020] The exception handling module is used to capture execution exceptions and trigger graph reconstruction.

[0021] Preferably, the reflective agent includes:

[0022] The process evaluation module is used to evaluate the efficiency and quality of the task execution process;

[0023] The result verification module is used to verify the correctness and completeness of the execution results;

[0024] The case generation module is used to encode the execution trajectory into structured cases;

[0025] The strategy optimization module is used to propose improvement suggestions based on the evaluation results.

[0026] Step S3: The agent executes sub-tasks based on the cognitive graph, feeds back the execution results to the cognitive graph, and uses the execution feedback to dynamically reconstruct the cognitive graph.

[0027] Step S4: The execution process is evaluated using an evolutionary mechanism based on case reasoning, structured cases are generated, and the structured cases are encoded and stored in the case library for reuse of experience in subsequent tasks.

[0028] Preferably, the evolutionary mechanism based on case-based reasoning includes:

[0029] The case study illustrates how historical tasks are encoded as quadruples containing a problem description, a solution path, execution feedback, and a success label.

[0030] Case encoding and storage: The problem description of the case is converted into a vector representation using an embedding model and stored in a vector database;

[0031] Case retrieval: When a new task is received, the description of the new task is converted into a vector, and a similarity search is performed in the case library;

[0032] Case reuse: Relevant cases retrieved are used as references to guide the construction of cognitive maps for new tasks;

[0033] Case updates: After the task is completed, the case library is updated based on the execution results, and new cases are directly added to the library; and some cases are merged or updated.

[0034] Preferably, the case library also includes a case update mechanism:

[0035] Once a new task is completed, add a success or failure label to the corresponding case based on the execution result;

[0036] For successful cases, extract key decision nodes and path patterns to enhance the reusability of the cases;

[0037] For failed cases, record the reasons for failure and the error path for reference in subsequent tasks;

[0038] Regularly organize the case library, merge similar cases, and delete outdated cases.

[0039] Step S5: The multi-agent cooperative network adopts a hybrid cooperative architecture that combines centralized and distributed approaches, and communicates between agents through a blackboard system.

[0040] Preferably, the hybrid collaborative architecture includes: a centralized architecture in the task planning phase, where the planning agent uniformly performs task decomposition and path planning; and a distributed architecture in the task execution phase, where multiple execution agents process different sub-tasks in parallel.

[0041] Preferably, communication using a blackboard system includes: agents communicating with each other through the blackboard system, which serves as a shared workspace for storing intermediate results and state information; and using an attention mechanism to filter redundant information and triggering high-bandwidth communication only at key decision nodes.

[0042] Step S6: Metacognitive monitoring is performed by monitoring the confidence level of the reasoning process in real time, and when multiple agents reach contradictory conclusions, an arbitration mechanism is used to arbitrate the conflict.

[0043] Preferably, the metacognitive monitoring mechanism includes: real-time monitoring of the confidence level of the current reasoning process; automatically triggering subtask splitting or calling a higher-capability expert agent when the confidence level is lower than a preset threshold; recording key decision points in the reasoning process, supporting path backtracking and correction, and dynamically adjusting the functional boundaries and prompting strategies of the agent according to the task complexity;

[0044] Preferably, the conflict arbitration mechanism includes: when multiple agents reach contradictory conclusions, introducing the role of a defender for critical testing, using a weighted voting method to arbitrate conflicting viewpoints, dynamically adjusting the weights based on the agents' historical accuracy, and recording the conflict resolution process as a case study for optimizing subsequent collaborative strategies.

[0045] Step S7: Once all target nodes in the cognitive graph are completed, summarize and output the execution results of each subtask, and then evaluate the system performance.

[0046] Preferably, the following metrics are used to evaluate system performance: task success rate, path efficiency, learning gain, and fault recovery rate.

[0047] On the other hand, the present invention also provides a self-evolving multi-agent cooperative system based on a dynamic cognitive graph, for executing the above-mentioned self-evolving multi-agent cooperative method based on a dynamic cognitive graph, comprising the following functional modules:

[0048] Dynamic Cognitive Graph Module: Used to build and maintain cognitive graphs in the form of directed acyclic graphs, supporting dynamic addition and deletion operations of nodes and edges;

[0049] Planning agent module: used to receive task input, perform task parsing, decomposition and path planning;

[0050] The execution agent module is used to execute sub-tasks based on the cognitive graph and provide feedback on the execution results.

[0051] Reflective Agent Module: Used to evaluate the execution process and generate structured cases;

[0052] Case Library Module: Used to store and retrieve historical task cases;

[0053] Collaborative communication module: used for information sharing and state synchronization among intelligent agents.

[0054] On the other hand, the present invention also provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements a self-evolving multi-agent cooperative method based on a dynamic cognitive graph.

[0055] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0056] (1) By introducing dynamic cognitive graph technology, this invention upgrades the multi-agent collaboration mode from "linear chain execution" to "dynamic graph structure evolution", endowing the system with metacognitive monitoring and path correction capabilities similar to human System 2. The system can reconstruct the reasoning path in real time based on execution feedback during runtime, breaking through the limitations of traditional static workflow and effectively solving the problem of "cognitive rigidity".

[0057] (2) This invention encodes the "problem-path-feedback" of historical tasks into structured cases through an evolutionary mechanism based on case-based reasoning, thus achieving a leap from "pre-trained knowledge retrieval" to "process experience reuse". The system possesses human-like online lifelong learning capabilities. As the number of tasks increases, the system entropy decreases, decision-making efficiency is significantly improved, and the "experience gap" problem is effectively solved.

[0058] (3) This invention adopts a hybrid collaborative architecture, which ensures the global optimality of decision-making at the macro planning level and the flexibility and robustness of operation at the micro execution level. By combining the blackboard system and the attention mechanism, efficient information sharing and redundancy filtering are achieved, reducing the communication cost of multi-agent collaboration.

[0059] (4) By combining multi-agent debate with dynamic programming, this invention can effectively suppress the illusion phenomenon of large language models and provide a logically interpretable decision basis. It has important application value for scenarios with high fault tolerance and extremely low fault tolerance, such as financial risk control, scientific research, and complex software engineering. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a schematic diagram of the overall architecture provided in an embodiment of the present invention;

[0062] Figure 2 This is a schematic diagram of the dynamic cognitive graph structure and reconstruction process provided in an embodiment of the present invention;

[0063] Figure 3 This is a schematic diagram of the multi-agent collaboration process provided in an embodiment of the present invention;

[0064] Figure 4 This is a schematic diagram of the evolutionary mechanism based on case-based reasoning provided in an embodiment of the present invention;

[0065] Figure 5 This is a schematic diagram of the hybrid collaboration architecture provided in an embodiment of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Example 1

[0068] In this embodiment, combined with Figure 1 As shown, the present invention provides a technical solution: a self-evolving multi-agent cooperation method based on a dynamic cognitive graph, comprising the following operational steps:

[0069] Step S1: Construct a dynamic cognitive graph, modeling the reasoning process as a directed acyclic graph. Nodes in the directed acyclic graph represent reasoning states or subtasks, and edges represent state transitions or dependencies.

[0070] In this embodiment, the specific implementation process includes:

[0071] Step S11: Define the dynamic cognitive graph structure.

[0072] For example, a Dynamic Cognitive Graph (DCG) is defined as a directed acyclic graph G=(V,E), where V is the set of nodes and E is the set of edges. Each node v∈V represents a reasoning state or subtask, containing the following attributes: node identifier (id), task description, execution status (pending / in execution / completed / failed), execution result (result), and confidence. Each edge e∈E represents a dependency relationship or state transition between nodes, containing the following attributes: source node (source), target node (target), edge type (type) (dependency / transition / alternative), and weight (weight).

[0073] Step S12: Initialize the dynamic cognitive map.

[0074] For example, the task description "Task" received from the user is parsed and decomposed by the planning agent. The chain-of-thought method is used to recursively decompose the macro-task into sub-problem trees. These sub-problem trees are then converted into directed acyclic graphs to establish dependencies between sub-tasks, generating an initial cognitive graph.

[0075] Step S13: Perform dynamic reconstruction of the dynamic cognitive map.

[0076] For example, such as Figure 2 As shown, during task execution, the graph is dynamically reconstructed based on execution feedback, including the following operations:

[0077] Node status update: When a subtask is completed, update the status and result of the corresponding node;

[0078] Path expansion: When a new sub-problem is discovered during execution, a new node is dynamically inserted into the graph, and a dependency relationship with the existing node is established;

[0079] Path backtracking: When a subtask fails, a backtracking mechanism is triggered to add alternative path nodes to the graph and explore alternative solutions;

[0080] Graph pruning: When redundant or invalid paths are detected, the graph is pruned and optimized by deleting the corresponding nodes and edges to improve the graph structure.

[0081] Step S2: Initialize a multi-agent cooperative network including a planning agent, an execution agent, and a reflective agent.

[0082] In this embodiment, the specific implementation process includes:

[0083] Step S21: The planning agent in the multi-agent cooperative network is responsible for the macro-level decision-making of the system. The planning agent is the builder and maintainer of the cognitive graph and also includes the following functional modules:

[0084] Task parsing module: Uses a large language model to understand user input and extract task objectives, constraints, and evaluation criteria;

[0085] Task decomposition module: Employs a recursive decomposition strategy to break down complex tasks into executable atomic subtasks;

[0086] Graph generation module: Organizes subtasks into a directed acyclic graph, establishes dependencies and execution priorities;

[0087] Path planning module: Plans the optimal execution path based on the graph structure, supporting hybrid scheduling of parallel and serial execution.

[0088] Step S22: Utilize the Executor Agent in the multi-agent cooperative network to execute specific sub-tasks. The Executor Agent serves as the interface between the system and the external environment, and includes the following functional modules:

[0089] Tool Invocation Module: Encapsulates the invocation interfaces of various external tools, including search engines, code executors, API calls, etc.

[0090] Information retrieval module: Retrieves the information required for the task from external sources such as knowledge bases, databases, and the internet;

[0091] Status reporting module: Provides real-time feedback on execution progress and intermediate results to the cognitive graph and blackboard system;

[0092] Exception handling module: Captures exceptions during execution and triggers corresponding error recovery or graph reconstruction operations.

[0093] Step S23: Utilize the reflector agent in the multi-agent cooperative network to supervise and evaluate the execution process. The reflector agent is the core of the system's self-optimization. It includes the following functional modules:

[0094] Process evaluation module: Monitors the efficiency indicators of task execution, including time consumption, resource usage, path length, etc.

[0095] Result verification module: Verifies the correctness of the execution results and detects potential errors and illusions;

[0096] Case generation module: Encodes the complete execution trajectory into structured cases, extracting key decision points and patterns;

[0097] Strategy optimization module: Analyzes the reasons for failure, proposes improvement suggestions, and optimizes subsequent execution strategies.

[0098] Step S3: The agent executes sub-tasks based on the cognitive graph, feeds back the execution results to the cognitive graph, and uses the execution feedback to dynamically reconstruct the cognitive graph.

[0099] In this embodiment, as Figure 3 As shown, the task execution process follows this flow:

[0100] Step S31: Select the currently executable node from the cognitive graph (the node where all prerequisites have been completed).

[0101] Step S32: Assign the selected nodes to the corresponding execution agents;

[0102] Step S33: The agent invokes tools to complete the sub-task and feeds the results back to the cognitive graph;

[0103] Step S34: The planning agent updates the graph state based on feedback, and triggers graph reconstruction if necessary;

[0104] Step S35: The intelligent agent reflects on the real-time monitoring and execution process and intervenes when an anomaly is detected;

[0105] Step S36: Repeat the above steps until all target nodes are completed or the termination condition is met.

[0106] Step S4: The execution process is evaluated using an evolutionary mechanism based on case reasoning, generating structured cases containing questions, paths, and feedback. These structured cases are then encoded and stored in a case library for reuse of experience in subsequent tasks.

[0107] In this embodiment, as Figure 4 As shown, the evolutionary mechanism of case-based reasoning includes the following steps:

[0108] Case representation: Each case is defined as a quadruple Case=(P,S,F,L), where P is the problem description, S is the solution path (the execution trajectory of the cognitive map), F is the execution feedback (including intermediate results and final results), and L is the success label (success / failure / partial success).

[0109] Case encoding and storage: An embedding model (such as Sentence-BERT) is used to convert the problem description P of the case into a vector representation, which is then stored in a vector database. Simultaneously, complete cases are stored in a structured format in a relational database, supporting precise retrieval and statistical analysis.

[0110] Case retrieval: When a new task is received, its description is converted into a vector, and a similarity search is performed in the case database. The retrieval strategy employs a weighted combination of semantic and structural similarity, prioritizing the return of successful cases. The similarity calculation formula is:

[0111]

[0112] Where Q is the new task, C is the historical cases, and α and β are weight parameters.

[0113] Case reuse: Relevant cases are used as references to guide the construction of cognitive maps for new tasks. For successful cases, the structure of their solution paths is reused; for failed cases, their flawed paths are avoided. Reuse strategies include:

[0114] Path template reuse: Use the graph structure of successful cases as a template to adapt to new tasks;

[0115] Decision-making reference: At critical decision-making junctures, refer to the decision-making choices in historical cases;

[0116] Error avoidance: Identify erroneous paths in historical failure cases and proactively avoid them during planning.

[0117] Case Updates: After a task is completed, the case library is updated based on the execution results. New cases are added directly to the library; for cases highly similar to existing cases, they are merged or updated. The case library is periodically cleaned up, deleting outdated cases and merging duplicate cases.

[0118] Step S5: The multi-agent cooperative network adopts a hybrid cooperative architecture that combines centralized and distributed approaches, with agents communicating through a blackboard system.

[0119] In this embodiment, as Figure 5 As shown, multi-agent collaboration employs a hybrid architecture:

[0120] Centralized planning: During the task planning phase, the planning agent is responsible for task decomposition and resource scheduling to ensure global optimality. The planning agent maintains the global cognitive graph and coordinates the work allocation among the executing agents.

[0121] Distributed execution: During the task execution phase, multiple execution agents work in parallel, each responsible for different sub-tasks within the cognitive graph. These agents possess a degree of autonomy, allowing them to make decisions and adjustments within a local scope.

[0122] Blackboard System Communication: Agents communicate with each other through the Blackboard System. The Blackboard System serves as a shared workspace and stores the following information:

[0123] Global status: current task progress, status of each agent, and resource usage;

[0124] Intermediate results: The execution results of each subtask, available for querying and use by other intelligent agents;

[0125] Message queue: Communication messages between agents, including requests, responses, and notifications.

[0126] Attention Filtering: To reduce communication overhead, an attention mechanism is introduced to filter redundant information. Each agent selectively focuses on relevant information in the blackboard system based on the current task requirements, rather than receiving all updates. High-bandwidth communication is triggered only at critical decision nodes (such as path forks or execution failures).

[0127] Step S6: Metacognitive monitoring is performed by monitoring the confidence level of the reasoning process in real time, and when multiple agents reach contradictory conclusions, an arbitration mechanism including a defense attorney mechanism, weighted voting, and evidence tracing is used to arbitrate the conflict.

[0128] In this embodiment, in metacognitive monitoring: the system monitors the confidence level of the reasoning process in real time, and the confidence level is calculated based on the following factors:

[0129] Execution success rate: The percentage of recent subtasks that were successfully executed;

[0130] Path consistency: The similarity between the current path and historical successful cases;

[0131] Result validation rate: The percentage of execution results that pass validation.

[0132] When the confidence level is lower than the preset threshold When this happens, the following responses are triggered: task splitting (further decomposing the current subtask), expert invocation (invoking a higher-capability expert agent), and path backtracking (returning to the previous decision node to make a new selection).

[0133] Conflict Arbitration: When multiple agents reach conflicting conclusions, the following arbitration mechanism is adopted:

[0134] Defender mechanism: Introducing specialized critical agents to question and test mainstream viewpoints;

[0135] Weighted voting: Weights are assigned based on the agent's historical accuracy, and weighted voting is used for decision-making.

[0136] Evidence tracing: requires all parties to provide the basis for their decisions and makes judgments based on the quality of the evidence.

[0137] Step S7: Once all target nodes in the cognitive graph are completed, summarize the results of each subtask, generate the final output, and evaluate the system performance.

[0138] In this embodiment, the specific implementation process includes:

[0139] Output: Once all target nodes in the cognitive graph are completed, the results of each subtask are summarized to generate the final output. The output includes: task execution results, execution path description (interpretability), and confidence assessment.

[0140] System evaluation: The following metrics are used to evaluate system performance:

[0141] Task success rate: The percentage of tasks that are successfully completed;

[0142] Path efficiency: The ratio of the actual execution path length to the optimal path length;

[0143] Learning gain: The increase in success rate and efficiency as the number of tasks increases;

[0144] Fault recovery rate: The percentage of tasks that are successfully recovered from and completed after an anomaly occurs.

[0145] Example 2

[0146] In this embodiment, the present invention provides a self-evolving multi-agent cooperative system based on a dynamic cognitive graph, used to execute the above-described self-evolving multi-agent cooperative method based on a dynamic cognitive graph, including the following functional modules:

[0147] Dynamic Cognitive Graph Module: Used to build and maintain cognitive graphs in the form of directed acyclic graphs, supporting dynamic addition and deletion operations of nodes and edges;

[0148] Planning agent module: used to receive task input, perform task parsing, decomposition and path planning;

[0149] The execution agent module is used to execute sub-tasks based on the cognitive graph and provide feedback on the execution results.

[0150] Reflective Agent Module: Used to evaluate the execution process and generate structured cases;

[0151] Case Library Module: Implemented based on a vector database, used to store and retrieve historical task cases;

[0152] Collaborative communication module: including blackboard system, used for information sharing and state synchronization among agents.

[0153] Example 3

[0154] In this embodiment, the present invention also provides a computing device, including a memory and a processor. The memory stores executable code, and when the processor executes the executable code, it implements a self-evolving multi-agent cooperative method based on a dynamic cognitive graph.

[0155] Example 4

[0156] In this embodiment, to further verify the effectiveness of the present invention, experimental comparisons were conducted on the GAIA benchmark dataset. GAIA includes three difficulty levels: Level 1, Level 2, and Level 3, with Level 3 involving multi-step reasoning and complex tool combinations. The experimental results are shown in Table 1.

[0157] Table 1 Comparison of different methods under the GAIA benchmark dataset

[0158] method Level 1 accuracy Level 2 accuracy Level 3 accuracy Single LLM (GPT-4) 45.2% 32.1% 18.5% AutoGen 52.3% 38.7% 24.2% MetaGPT 54.1% 40.5% 26.8% Method of the present invention 61.8% 49.2% 35.6%

[0159] Experimental results show that the method proposed in this invention is superior to the comparative methods at all difficulty levels, especially in the Level 3 high-difficulty task, which verifies the effectiveness of dynamic cognitive graph and case reasoning mechanism in handling complex long-link reasoning tasks.

[0160] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0161] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A self-evolving multi-agent cooperative method based on dynamic cognitive graphs, characterized in that: The following steps are included: Step S1: Construct a dynamic cognitive graph, modeling the reasoning process as a directed acyclic graph; Step S2: Initialize a multi-agent cooperative network including a planning agent, an executive agent, and a reflective agent; Step S3: The agent executes sub-tasks based on the cognitive graph, feeds back the execution results to the cognitive graph, and uses the execution feedback to dynamically reconstruct the cognitive graph; Step S4: The execution process is evaluated using an evolutionary mechanism based on case reasoning, structured cases are generated, and the structured cases are encoded and stored in the case library for reuse of experience in subsequent tasks. Step S5: The multi-agent cooperative network adopts a hybrid cooperative architecture that combines centralized and distributed approaches, and communicates between agents through a blackboard system. Step S6: Metacognitive monitoring is performed by monitoring the confidence level of the reasoning process in real time, and when multiple agents reach contradictory conclusions, an arbitration mechanism is used to arbitrate the conflict. Step S7: Once all target nodes in the cognitive graph are completed, summarize and output the execution results of each subtask, and then evaluate the system performance.

2. The self-evolving multi-agent cooperation method based on dynamic cognitive graphs according to claim 1, characterized in that: In step S1, nodes in the directed acyclic graph represent reasoning states or subtasks, and edges represent state transitions or dependencies. Constructing the dynamic cognitive graph includes: Define the structure of a dynamic cognitive graph; Receive the task description (Task) input by the user, and use a planning agent to parse and decompose the task. During task execution, the cognitive map is dynamically reconstructed based on the execution feedback.

3. The self-evolving multi-agent cooperation method based on dynamic cognitive graphs according to claim 2, characterized in that: Step S13, which involves dynamically reconstructing the cognitive map based on execution feedback, includes: When a subtask is executed successfully, the corresponding node is marked as completed and the successor node is activated. When a subtask fails, a backtracking mechanism is triggered, adding alternative path nodes to the cognitive graph. When new sub-problems are discovered during execution, new nodes are dynamically inserted into the cognitive graph, and dependencies on existing nodes are established. When redundant or invalid paths are detected, the cognitive graph is pruned and optimized by deleting the corresponding nodes and edges.

4. The self-evolving multi-agent cooperation method based on dynamic cognitive graphs according to claim 3, characterized in that: The planning of the intelligent agent in step S2 includes: Task parsing module: Used to understand user input using a large language model, and extract task objectives, constraints and evaluation criteria; Task decomposition module: Used to decompose complex tasks into executable atomic subtasks using a recursive decomposition strategy; Graph generation module: used to organize subtasks into a directed acyclic graph, establish dependencies and execution priorities; Path planning module: Used to plan the optimal execution path based on the graph structure.

5. The self-evolving multi-agent cooperation method based on dynamic cognitive graphs according to claim 4, characterized in that: The execution of the agent in step S2 includes: Tool Invocation Module: Used to encapsulate the invocation interface of external tools; Information retrieval module: used to obtain the information required for the task from external data sources; Status reporting module: used to provide real-time feedback on execution progress and intermediate results to the cognitive graph and blackboard system; Exception handling module: Used to capture exceptions during execution and trigger corresponding error recovery or graph reconstruction operations.

6. The self-evolving multi-agent cooperation method and system based on dynamic cognitive graphs according to claim 5, characterized in that: The reflective agent in step S2 includes: Process evaluation module: Used to monitor efficiency metrics for task execution; Result verification module: Used to verify the correctness of the execution results; Case generation module: used to encode the complete execution trajectory into structured cases and extract key decision points and patterns; Strategy optimization module: Used to analyze the reasons for failure, propose improvement suggestions, and optimize subsequent execution strategies.

7. The self-evolving multi-agent cooperation method based on dynamic cognitive graphs according to claim 6, characterized in that: The evolutionary mechanism based on case-based reasoning in step S4 includes: The case study illustrates how historical tasks are encoded as quadruples containing a problem description, a solution path, execution feedback, and a success label. Case encoding and storage: The problem description of the case is converted into a vector representation using an embedding model and stored in a vector database; Case retrieval: When a new task is received, the description of the new task is converted into a vector, and a similarity search is performed in the case library; Case reuse: Relevant cases retrieved are used as references to guide the construction of cognitive maps for new tasks; Case updates: After the task is completed, the case library is updated based on the execution results, and new cases are directly added to the library; and some cases are merged or updated.

8. The self-evolving multi-agent cooperation method based on dynamic cognitive graphs according to claim 7, characterized in that: Step S5 includes: A centralized architecture is adopted in the task planning phase, with the planning agent uniformly performing task decomposition and path planning; A distributed architecture is adopted during the task execution phase, with multiple execution agents processing different sub-tasks in parallel. Intelligent agents communicate with each other through a blackboard system, which serves as a shared workspace for storing intermediate results and status information. An attention mechanism is used to filter redundant information, triggering high-bandwidth communication only at key decision nodes.

9. The self-evolving multi-agent cooperation method based on dynamic cognitive graphs according to claim 8, characterized in that: Step S6 includes: Metacognitive monitoring mechanism: Real-time monitoring of the confidence level of the current reasoning process; When the confidence level is lower than a preset threshold, it automatically triggers subtask splitting or calls a higher-capability expert agent; Records key decision points in the reasoning process, supports path backtracking and correction, and dynamically adjusts the functional boundaries and prompting strategies of the agent according to the task complexity; Conflict arbitration mechanism: When multiple agents reach contradictory conclusions, a defender role is introduced to conduct critical testing. A weighted voting method is used to arbitrate conflicting viewpoints. The weights are dynamically adjusted based on the agents' historical accuracy. The conflict resolution process is recorded as a case study to optimize subsequent collaborative strategies.

10. A self-evolving multi-agent cooperative system based on a dynamic cognitive graph, characterized in that: The self-evolving multi-agent cooperative method based on dynamic cognitive graphs as described in any one of claims 1-9 includes: Dynamic Cognitive Graph Module: Used to construct and maintain cognitive graphs in the form of directed acyclic graphs; Planning agent module: used to receive task input, perform task parsing, decomposition and path planning; The execution agent module is used to execute sub-tasks based on the cognitive graph and provide feedback on the execution results. Reflective Agent Module: Used to evaluate the execution process and generate structured cases; Case Library Module: Used to store and retrieve historical task cases; Collaborative communication module: used for information sharing and state synchronization among intelligent agents.

Images