Multi-agent cooperation enhancement method, system and equipment based on knowledge graph

Abstract

The invention discloses a multi-agent cooperation enhancement method, system and equipment based on a knowledge graph, and the method comprises the steps: obtaining original data in an external environment, carrying out the preprocessing and feature extraction of the original data, generating a knowledge triple, storing the knowledge triple in a local knowledge graph, and submitting the knowledge triple to a shared knowledge graph for knowledge updating; when a to-be-executed task is received, decomposing the to-be-executed task by utilizing the large language model and querying global knowledge in the shared knowledge graph and local knowledge in the local knowledge graph to obtain a plurality of sub-tasks; a bipartite graph minimum cost matching algorithm is adopted to match a plurality of sub-tasks with the capability and availability of each agent to generate a preliminary task allocation scheme, and a large language model is utilized to optimize the preliminary task allocation scheme to generate an optimal task allocation scheme; and sending each task allocation knowledge fragment in the optimal task allocation scheme to a corresponding agent for collaborative execution through a semantic communication protocol.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a multi-agent collaboration enhancement method, system and device based on a knowledge graph. BACKGROUND

[0002] In the application expansion of multi-agent systems (MAS), although it has shown significant flexibility and self-organizing advantages in complex problem fields such as intelligent manufacturing and smart cities. However, despite the great potential of multi-agent systems, current systems still face many challenges, the core reason being the difficulty in bridging the "knowledge gap". Including the information island phenomenon leading to low efficiency of knowledge sharing between agents; the limitations of traditional communication mechanisms in deep semantic transmission make it difficult for agents to build a unified global knowledge view, often limited to local decision-making, and it is difficult to reach the optimal global solution when collaborating. Although the knowledge graph provides a powerful tool for semantic knowledge organization, existing explorations of integrating it with MAS are mostly limited to assisting decision-making or single knowledge updating, lacking a systematic collaboration mechanism and knowledge dynamic evolution logic. Especially in the multi-agent collaboration framework, although there are role allocation and dialogue mechanisms based on large language models, they have not deeply integrated knowledge graphs to form a self-enhancing closed-loop collaboration system.

[0003] Therefore, there is an urgent need for a method that can efficiently utilize the semantic organization advantages of knowledge graphs and deeply integrate the dynamic collaboration needs of multi-agent systems, achieving collaborative optimization of knowledge sharing, task decomposition, and agent capability matching, breaking through the limitations of existing MAS in complex task execution, and improving the overall collaboration efficiency of the system. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a multi-agent collaboration enhancement method, device, system and equipment based on a knowledge graph, aiming to solve the problem of how to optimize the knowledge sharing mechanism and task allocation collaboration of multi-agent systems in a dynamic environment, thereby enhancing the overall collaboration capability of multi-agent systems.

[0005] To achieve the above-mentioned purpose, the present application provides a multi-agent collaboration enhancement method based on a knowledge graph, which comprises: Obtaining raw data in an external environment, preprocessing and feature extraction of the raw data, generating knowledge triples and storing them in a local knowledge graph, and submitting the knowledge triples to a shared knowledge graph in an incremental manner based on a semantic communication protocol for knowledge updating; When receiving a task to be executed, using a large language model and querying global knowledge in the shared knowledge graph and local knowledge in the local knowledge graph to decompose the task to be executed, obtaining a plurality of sub-tasks; A bipartite graph minimum cost matching algorithm is used to match multiple subtasks with the capabilities and availability of each agent to generate a preliminary task allocation scheme. The preliminary task allocation scheme is then optimized using a large language model to generate the optimal task allocation scheme. Each task allocation knowledge fragment in the optimal task allocation scheme is sent to the corresponding intelligent agent for collaborative execution via a semantic communication protocol.

[0006] Preferably, the step of submitting the knowledge triples to the shared knowledge graph incrementally for knowledge updating based on the semantic communication protocol includes: When semantic conflicts are detected between knowledge triples submitted by different agents, arbitration is performed by activating a conflict resolution agent based on a large language model, and the knowledge triples that meet the conditions are used as knowledge to update the shared knowledge graph after arbitration.

[0007] Preferably, the arbitration by activating a conflict resolution agent based on a large language model includes: Send structured debate prompts to agents with semantic conflicts and collect evidence and reasoning paths provided by each agent; Arbitration of evidence and reasoning paths is based on large-scale language models by combining the credibility and confidence of information sources.

[0008] Preferably, the step of using a bipartite graph minimum cost matching algorithm to match multiple subtasks with the capabilities and availability of each agent to generate a preliminary task allocation scheme includes: Construct a bipartite graph with the set of subtasks and the set of agents as vertices; The matching cost weights of each edge in the bipartite graph are calculated based on the capability attributes, state attributes, and historical performance data of each agent recorded in the shared knowledge graph. The optimal subtask-agent allocation pair is solved by the minimum cost matching algorithm to generate the preliminary task allocation scheme.

[0009] Preferably, the step of optimizing the initial task allocation scheme using a large language model to generate an optimal task allocation scheme includes: The preliminary task allocation scheme, task context description, and agent metadata extracted from the shared knowledge graph are input into a large language model. The initial allocation scheme is evaluated and adjusted using a semantic bidding mechanism based on a contract network protocol using a large language model, generating interpretable reasons for task allocation and the optimal task allocation scheme.

[0010] Preferably, the step of sending each task allocation knowledge fragment in the optimal task allocation scheme to the corresponding intelligent agent for collaborative execution via a semantic communication protocol includes: Based on the task-assigned knowledge fragment, query the shared knowledge graph to see if there is an intent conflict in the intent operation; If an intent conflict is detected, a large language model combined with background knowledge from a shared knowledge graph is used to arbitrate and generate a conflict resolution scheme, so that the conflicting agents can perform collaborative execution according to the conflict resolution scheme.

[0011] Preferably, the semantic communication protocol supports multi-level transmission modes, including: In bandwidth-constrained scenarios, the sender transmits semantic packets consisting of embedded representations of key entities in the knowledge graph subgraph, and the receiver performs semantic reconstruction based on the local knowledge graph and context description. In scenarios with high semantic consistency, the sender transmits a semantic packet consisting of structured triples and embedded representations of knowledge graph subgraphs. The receiver then uses embedding clustering, entity linking, knowledge graph completion, and context alignment strategies to reconstruct the structure of knowledge fragments and fuse the graph.

[0012] To achieve the above objectives, the present invention also provides a knowledge graph-based multi-agent collaboration enhancement device, the device comprising: The data acquisition unit is used to acquire raw data from the external environment, preprocess and extract features from the raw data, generate knowledge triples and store them in the local knowledge graph, and submit the knowledge triples to the shared knowledge graph in an incremental manner based on the semantic communication protocol for knowledge updates. The task decomposition unit is used to decompose the task to be executed into multiple sub-tasks when a task to be executed is received, by using a large language model and querying global knowledge in the shared knowledge graph and local knowledge in the local knowledge graph. The scheme generation unit is used to match multiple subtasks with the capabilities and availability of each agent using a bipartite graph minimum cost matching algorithm to generate a preliminary task allocation scheme, and to optimize the preliminary task allocation scheme using a large language model to generate an optimal task allocation scheme. The collaborative execution unit is used to send each task allocation knowledge fragment in the optimal task allocation scheme to the corresponding intelligent agent for collaborative execution through a semantic communication protocol.

[0013] To achieve the above objectives, the present invention also provides a multi-agent system, the system comprising: The perception module is used to acquire raw data from the external environment and send it to the local knowledge graph module; The local knowledge graph module is used to preprocess and extract features from raw data, generate knowledge triples as local knowledge of the agent and store them, and perform knowledge reasoning and context understanding according to the needs of the decision module. The shared knowledge graph module is used to receive knowledge update requests from various agents, realize knowledge updates by incrementally constructing knowledge triples, and receive knowledge query requests from various agents to provide global knowledge. The decision-making module is used to obtain local knowledge from the local knowledge graph module and global knowledge from the shared knowledge graph module to perform task decomposition and generate execution instructions for task allocation schemes. The communication module is used to encapsulate execution instructions into corresponding knowledge fragments and send them to the corresponding intelligent agent through a semantic communication protocol, as well as to submit knowledge triples to the shared knowledge graph.

[0014] To achieve the above objectives, the present invention also proposes a knowledge graph-based multi-agent collaboration enhancement device, comprising a processor, a memory, and a computer program stored in the memory, wherein the computer program is executed by the processor to implement the steps of a knowledge graph-based multi-agent collaboration enhancement method as described in the above embodiments.

[0015] To achieve the above objectives, the present invention also proposes a computer-readable storage medium storing a computer program that is executed by a processor to implement the steps of a knowledge graph-based multi-agent collaboration enhancement method as described in the above embodiments.

[0016] To achieve the above objectives, the present invention also proposes a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of a knowledge graph-based multi-agent collaboration enhancement method as described in the above embodiments.

[0017] Beneficial effects: The above scheme acquires raw data from the environment, performs preprocessing and feature extraction to generate knowledge triples, and submits incremental knowledge triples to a shared knowledge graph based on a semantic communication protocol. This shared knowledge graph provides all agents with a continuously updated and consistent global situation, fundamentally solving the problem of existing systems struggling to cope with sudden environmental changes due to information silos and local decision-making. This achieves efficient knowledge sharing and updating among agents, improving the overall cognitive level of the system. During task execution, a large-scale language model combined with global and local knowledge is used for task decomposition and allocation, ensuring the scientific and rational nature of task allocation and fully utilizing the capabilities of each agent. This significantly improves the task execution efficiency and overall collaboration capabilities of the multi-agent system in dynamic environments, enhancing the system's intelligence and adaptability.

[0018] When semantic conflicts are detected in knowledge triples submitted by different agents, a conflict resolution agent based on a large language model is activated for arbitration. This effectively solves the problem of knowledge inconsistency in multi-agent systems, ensuring the accuracy and reliability of the shared knowledge graph. By issuing structured debate prompts and collecting evidence and reasoning paths provided by each agent for arbitration, the "debate" between agents encourages them to analyze and verify their own knowledge more deeply, improving the quality and credibility of the knowledge and thus enhancing the system's transparency and interpretability. Furthermore, the system can promptly correct erroneous information in dynamic environments, avoiding decision-making errors and collaborative chaos caused by knowledge conflicts, further improving the system's stability and collaborative efficiency.

[0019] A bipartite graph minimum cost matching algorithm is used to match multiple subtasks with the capabilities and availability of each agent, generating a preliminary task allocation scheme. This algorithm fully considers the individual differences of agents and the characteristics of tasks, achieving precise matching between tasks and agent capabilities. By constructing a bipartite graph with the set of subtasks and the set of agents as vertices, and calculating the matching cost weight based on the agent capability attributes, state attributes, and historical performance data recorded in the shared knowledge graph, the algorithm ensures the rationality and efficiency of task allocation, maximizes the advantages of each agent, and improves the overall performance of the system and the quality of task completion.

[0020] By utilizing a large-scale language model to optimize the initial task allocation scheme and generate the optimal scheme, the scientific rigor and rationality of task allocation are further enhanced. The large-scale language model evaluates and adjusts the initial allocation scheme based on a semantic bidding mechanism using a contract network protocol, fully tapping the potential of the agents and discovering better task allocation combinations. Simultaneously, generating interpretable reasons for task allocation helps agents understand and execute the assigned tasks, enhancing the system's operability and transparency. This enables multi-agent systems to better coordinate and cooperate in complex tasks, improving the success rate and efficiency of task completion.

[0021] In the process of sending task allocation knowledge fragments to corresponding agents for collaborative execution via semantic communication protocols, the system queries the shared knowledge graph for intent conflicts and uses a large language model combined with background knowledge from the shared knowledge graph to arbitrate and generate conflict-free decision schemes. This effectively avoids collaboration failures or resource waste caused by intent conflicts during task execution. The system can adjust and optimize task execution schemes in a timely manner, ensuring the consistency and coherence of collaborative work among agents, improving the system's robustness and task execution reliability, and enabling the multi-agent system to have stronger adaptability and collaborative capabilities in dynamic environments.

[0022] The semantic communication protocol supports multi-level transmission modes. By transmitting refined knowledge graph subgraphs, it significantly reduces the amount of communication data and unnecessary computational overhead. It flexibly selects transmission methods according to different scenario requirements. In bandwidth-constrained scenarios, it transmits semantic packets composed of embedded representations of key entities in the knowledge graph subgraph. The receiver then performs semantic reconstruction based on its local knowledge graph and context description, which effectively reduces the amount of communication data, improves communication efficiency, and ensures that information interaction between agents can still be carried out quickly and accurately under limited bandwidth conditions. In scenarios with high semantic consistency, the sender transmits semantic packets composed of embedded representations of structured triples and knowledge graph subgraphs. The receiver then performs structural reconstruction and graph fusion of knowledge fragments through embedding clustering, entity linking, knowledge graph completion, and context alignment strategies, ensuring the integrity and accuracy of knowledge transmission, enhancing the system's semantic understanding and knowledge integration capabilities, and further improving the collaborative effect and task execution quality of multi-agent systems, enabling them to better adapt to complex and dynamic environments. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a knowledge graph-based multi-agent collaboration enhancement method according to an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of a multi-agent system provided in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of path planning based on knowledge graph enhancement, provided as an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the structure of a knowledge graph-based multi-agent collaboration enhancement device according to an embodiment of the present invention.

[0028] The realization of the invention's objective, its functional characteristics, and advantages will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 a part of the embodiments of the present invention, not all of them. 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. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. 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.

[0030] The present invention will be described in detail below with reference to the embodiments.

[0031] Reference Figure 1 The diagram shown is a flowchart of a knowledge graph-based multi-agent collaboration enhancement method provided in an embodiment of the present invention.

[0032] In this embodiment, the method is implemented based on a multi-agent system; wherein, the system includes: The perception module is used to acquire raw data from the external environment and send it to the local knowledge graph module; The local knowledge graph module is used to preprocess and extract features from raw data, generate knowledge triples as local knowledge of the agent and store them, and perform knowledge reasoning and context understanding according to the needs of the decision module. The shared knowledge graph module is used to receive knowledge update requests from various agents, realize knowledge updates by incrementally constructing knowledge triples, and receive knowledge query requests from various agents to provide global knowledge. Furthermore, the shared knowledge graph module is also used to arbitrate by activating a conflict resolution agent based on a large language model when semantic conflicts are detected between knowledge triples submitted by different agents. The decision-making module is used to obtain local knowledge from the local knowledge graph module and global knowledge from the shared knowledge graph module to perform task decomposition and generate execution instructions for task allocation schemes. The communication module is used to encapsulate execution instructions into corresponding knowledge fragments and send them to the corresponding intelligent agent through a semantic communication protocol, as well as to submit knowledge triples to the shared knowledge graph.

[0033] Reference Figure 2As shown, the system architecture includes a perception module, a local knowledge graph module, a shared knowledge graph module, a decision-making module, and a communication module. The perception module is responsible for acquiring raw data (such as sensor readings, text information, images, and sounds) from the external environment. This is the system's initial input layer, providing raw data for knowledge extraction. It performs preliminary data preprocessing (such as noise reduction and formatting) and then transmits the processed data (such as events, objects, and relationships) to the local knowledge graph module.

[0034] The local knowledge graph module is a private memory maintained by each agent to store its own private, local knowledge graph. It stores the agent's own private knowledge, task-related context, and local knowledge. This knowledge is refined and updated through continuous learning and reasoning. The local knowledge graph helps agents to perform individual understanding, context-specific processing, and make quick, preliminary local decisions even without global knowledge information. It receives structured data from the perception module, transforms it into knowledge triples (such as subject, predicate, and object), and stores them. At the same time, the local knowledge graph module performs knowledge reasoning and context understanding according to the needs of the decision module, and then provides the decision module with refined local knowledge or knowledge fragments for decision-making. It also submits some newly discovered knowledge (such as new knowledge triples) to the shared knowledge graph module for knowledge updates.

[0035] The shared knowledge graph module is the core, global, and dynamically updated knowledge base of the entire system. It can be a centralized or distributed shared knowledge base. It is responsible for global knowledge management, dynamic evolution, conflict resolution, and multimodal fusion, providing all agents with a unified and consistent global situational view. Its processing includes: incrementally building knowledge graphs by receiving knowledge update requests from the local knowledge graph modules of each agent; arbitrating conflict resolution algorithms (such as LLM proxy arbitration) when multiple agents' update requests have semantic contradictions, ensuring the integrity and credibility of knowledge; receiving knowledge query requests from the decision modules of each agent, providing global knowledge to support decision-making; and then feeding back the update result (success or failure) to the agent requesting the update, or returning structured knowledge graph subgraph information to the agent requesting the query.

[0036] The decision-making module enables the agent to leverage the powerful reasoning capabilities of LLM (Locally Principles Management) to make optimal decisions, including task allocation, path planning, and conflict resolution. It acquires local knowledge from the local knowledge graph module and sends task-related knowledge queries to the shared knowledge graph module to obtain global knowledge. By combining local and global knowledge, it performs task decomposition, path planning, and conflict resolution to generate specific execution instructions, such as "move to location X" or "execute task Y." These execution instructions are then transmitted to the communication module for information synchronization and task collaboration with other agents.

[0037] The communication module uses a knowledge graph-based semantic communication protocol to transform communication content from raw data or text into structured knowledge graph subgraphs. It is responsible for transmitting and receiving knowledge fragments (e.g., knowledge graph subgraphs) instead of raw data, significantly improving communication efficiency and accuracy. It receives execution instructions and intentions from the decision-making module, encapsulates them into semantically clear knowledge fragments, sends semantic communication data to other agents, and simultaneously receives semantic communication data from other agents. It is also responsible for sending and receiving control signals (such as task allocation and confirmation information). The received knowledge fragments are transmitted to the local perception module or the local knowledge graph module to update their local knowledge or trigger new decisions.

[0038] The following example illustrates this: Suppose the task to be performed is to write a complex technical document, including product features, code examples, and user guides. The system consists of multiple LLM agents, including a task coordinator agent, a product agent, a development agent, a content agent, and an auditing agent. Each agent has a perception module, a local knowledge graph module, a shared knowledge graph module, a decision-making module, and a communication module.

[0039] The task coordinator agent's perception module acquires raw requirement data from project management tools (such as Jira and Confluence), and then transforms this raw data into initial knowledge for storage in a local knowledge graph, such as (document, hasTopic, product feature X) and (document, hasDeadline, 2025-09-30). Furthermore, the task coordinator agent incrementally updates the shared knowledge graph with this initial knowledge. When the system receives a task, the task coordinator agent intelligently decomposes the task into multiple executable sub-tasks using task relationships such as isSubTaskOf in the LLM and the shared knowledge graph, such as "writing feature descriptions," "writing code examples," and "creating user guides." It initiates semantic bidding to other LLM agents (product, development, content, and review agents). Each agent evaluates its capabilities and availability based on its own local knowledge graph and bids to the task coordinator agent. Then, the task coordinator agent uses a knowledge graph-based bipartite graph minimum cost matching algorithm to generate a preliminary task allocation scheme, and uses a large language model to optimize the preliminary task allocation scheme to generate the optimal task allocation scheme.

[0040] Furthermore, the task coordinator agent sends corresponding task assignment knowledge fragments (e.g., the content agent, assignedTask, writing a user guide) to other LLM agents via a semantic communication protocol for collaborative execution. For example, after receiving the corresponding subtask, the content agent sends a semantic query to the product agent to obtain information such as (function X, hasUserBenefit, benefit content); the product agent queries its local knowledge graph and returns it as a structured knowledge graph subgraph, rather than a lengthy requirements document; when writing code examples, the development agent submits refined knowledge triples to the shared knowledge graph, such as (function Z, hasExampleCode, code fragment).

[0041] When a developer agent updates a function and submits a new knowledge triple to the shared knowledge graph, the task coordinator agent continuously monitors the graph. If a contradiction is found between the new knowledge and existing knowledge (e.g., a draft document submitted by the content agent), the system immediately activates an LLM-based conflict resolution algorithm agent. This agent queries the knowledge source credibility in the graph and sends "debate prompts" to both conflicting parties. Based on the arbitration result, the task coordinator agent sends a correction instruction to the content agent, such as "Please update the parameter description of function Z in the document to be consistent with the latest codebase."

[0042] When the document writing task is nearing completion, the task coordinator agent notifies the review agent to begin work. The review agent's decision module queries the shared knowledge graph, comparing the document content with the facts, grammar rules, and style guidelines recorded in the shared knowledge graph. The review agent identifies and corrects errors (such as a spelling mistake). This correction action leaves a traceable reasoning path in the shared knowledge graph, for example: (correction action, basedOn, (grammar rule, isErrorOf, spelling mistake)). This provides a transparent and reliable guarantee for the final document quality.

[0043] Reference Figure 1 As shown, the method includes: S11, acquire raw data from the external environment, preprocess and extract features from the raw data, generate knowledge triples and store them in the local knowledge graph, and submit the knowledge triples to the shared knowledge graph in an incremental manner based on the semantic communication protocol for knowledge update.

[0044] Furthermore, in step S11, the step of incrementally submitting the knowledge triples to the shared knowledge graph for knowledge updating based on the semantic communication protocol includes: When semantic conflicts are detected between knowledge triples submitted by different agents, arbitration is performed by activating a conflict resolution agent based on a large language model, and the knowledge triples that meet the conditions are used as knowledge to update the shared knowledge graph after arbitration.

[0045] Furthermore, the arbitration process by activating a conflict resolution agent based on a large language model includes: Send structured debate prompts to agents with semantic conflicts and collect evidence and reasoning paths provided by each agent; Arbitration of evidence and reasoning paths is based on large-scale language models by combining the credibility and confidence of information sources.

[0046] In this embodiment, an agent acquires raw data from the external environment, performs preprocessing and feature extraction, generates knowledge triples, and stores them in a local knowledge graph. When the LLM-based knowledge graph completion framework (GS-KGC) detects new information, it submits the knowledge triples incrementally to the shared knowledge graph for knowledge updates based on a semantic communication protocol. For example, when an agent discovers a new entity or relation while performing a task, it can submit the new knowledge triple (subject, predicate, object) to the shared knowledge graph incrementally. To support this incremental construction, a multi-agent enhanced knowledge extraction and completion mechanism based on a Large Language Model (LLM) is utilized. This mechanism uses LLM as its core engine and can automatically extract entities and relations from unstructured or semi-structured data provided by agents, and mark new entities and relations for addition. This process uses a multi-agent collaborative model, with multiple LLM agents (such as entity extraction agents and relation extraction agents) working together to ensure the accuracy and breadth of the extraction. This shift from "data ingestion" to "knowledge management" means that maintaining a shared knowledge graph is a continuous and proactive process, constantly validating, refining, and expanding the knowledge graph.

[0047] When contradictory knowledge triples uploaded by different agents occur, the system activates a Conflict Resolution Algorithm Agent (CRA) based on LLM (Local Level Management) for arbitration. This CRA uses "debate prompts" to determine the consistency (e.g., agreement, contradiction, ambiguity) between newly extracted knowledge triples and existing knowledge, considering domain knowledge and partial contextual descriptions. It sends structured prompts to the conflicting agents, requesting them to provide evidence and reasons. High-confidence conflicts may be flagged for manual review, and older / lower-confidence data may be overwritten. For example, when agents A and B upload contradictory facts about the same entity, the CRA prompts them to "state your position on fact X and provide evidence and reasoning." Ultimately, the CRA arbitrates based on the evidence provided by the agents, confidence levels, the credibility of the information source, and the logical reasoning capabilities of LLM. This mechanism can handle unseen entities and complex logical inconsistencies, and allows setting a threshold for manual review intervention for high-confidence conflicts to ensure the integrity and credibility of the knowledge base.

[0048] Furthermore, the system maintains a dynamic memory structure for the shared knowledge graph, namely "known knowledge (Kt)" and "required knowledge (Rt)," forming a continuously refined and expanded closed-loop feedback. Kt continuously refines the shared knowledge graph through interaction with the external environment, while Rt guides the system to prioritize the acquisition of unresolved information, thereby avoiding redundancy and improving the efficiency of knowledge acquisition.

[0049] S12, when a task to be executed is received, the task to be executed is decomposed into multiple sub-tasks by using a large language model and querying global knowledge in the shared knowledge graph and local knowledge in the local knowledge graph.

[0050] S13. The bipartite graph minimum cost matching algorithm is used to match multiple subtasks with the capabilities and availability of each agent to generate a preliminary task allocation scheme. The preliminary task allocation scheme is then optimized using a large language model to generate the optimal task allocation scheme.

[0051] Furthermore, in step S13, the bipartite graph minimum cost matching algorithm is used to match multiple subtasks with the capabilities and availability of each agent to generate a preliminary task allocation scheme, including: S13-1, Construct a bipartite graph with the set of subtasks and the set of agents as vertices; S13-2, calculate the matching cost weight of each edge in the bipartite graph based on the ability attributes, state attributes and historical performance data of each agent recorded in the shared knowledge graph; S13-3, Solve for the optimal subtask-agent allocation pair using the minimum cost matching algorithm to generate the preliminary task allocation scheme.

[0052] Furthermore, in step S13, optimizing the preliminary task allocation scheme using a large language model to generate an optimal task allocation scheme includes: S13-4, Input the preliminary task allocation scheme, task context description and agent metadata extracted from the shared knowledge graph into the large language model; S13-5, using a large language model based on the semantic bidding mechanism of the contract network protocol, the preliminary allocation scheme is evaluated and adjusted to generate interpretable task allocation reasons and the optimal task allocation scheme.

[0053] In this embodiment, when the system receives a task, the agent first queries the shared knowledge graph and intelligently decomposes the task into smaller subtasks using task relationships / dependencies (such as "isSubTaskOf"). Subsequently, the system employs a knowledge graph-based bipartite graph minimum-cost matching algorithm (KG-enhanced Min-Cost BipartiteMatching) to match subtasks with agents based on the agent capabilities and availability recorded in the knowledge graph, generating a preliminary task allocation scheme to achieve efficient and intelligent resource allocation. This includes constructing a bipartite graph with the subtask set and the agent set as vertices, calculating the matching cost weight of each edge in the bipartite graph based on the capability attributes, state attributes, and historical performance data of each agent recorded in the shared knowledge graph, and solving for the optimal subtask-agent allocation pair using the minimum-cost matching algorithm to generate the preliminary task allocation scheme.

[0054] Then, using LLM (which serves as the core decision engine and leverages its powerful semantic understanding capabilities), a semantic bidding mechanism based on the Contract Network protocol is employed to analyze and optimize the initial task allocation scheme, generating the optimal task allocation scheme. LLM can generate an interpretable optimal task allocation scheme based on the task context description and metadata in the knowledge graph (such as "Agent B is good at syntax and style checking"), and explain the rationale behind the task allocation. For example, it might state: "It is recommended that Agent B prioritize proofreading the 'Introduction' written by Agent A, because the metadata in the shared knowledge graph indicates that Agent B is good at syntax and style checking, while Agent A has recently made significant modifications to this section, potentially leading to formatting errors." This approach achieves more interpretable and reasonable task allocation.

[0055] Furthermore, in collaborative tasks, before executing an operation, an agent queries the shared knowledge graph to check if its intended operation conflicts with other agents' current or anticipated operations. By employing a knowledge graph-based optimistic concurrency control (OCC-KG) mechanism, potential conflicts in the intended operation are quickly identified by comparing the agent's intended operation with the "lock information" in the shared knowledge graph in real time. Once an intention conflict is detected, the agent submits relevant conflict information (e.g., the editing location of the conflict, the conflicting agent ID, and their intentions) to the LLM. The LLM, combined with background knowledge in the shared knowledge graph (such as task priority, geographical location, traffic conditions, etc.), uses a meta-knowledge-driven arbitration algorithm to propose an optimal conflict resolution solution. This could be, for example, "suggesting agent A wait for agent B to complete its operation before proceeding," or "suggesting agent A modify another part of the document," or planning alternative paths for one of the agents. In path planning, this mechanism can be combined with existing conflict search algorithms (CBS algorithm) to predict and resolve potential collisions before execution, thereby improving the conflict-free success rate of global planning.

[0056] By sharing knowledge graphs, intelligent agents can achieve more advanced collaborative planning. For example, when collaboratively writing reports, agents can automatically adjust their work order based on dependencies in the shared knowledge graph, using the KG-based Dependency Management and Task Sequencing (KG-DMTS) algorithm. This ensures that subsequent tasks are executed only after a prerequisite task is completed, thus avoiding unnecessary rework and logical errors. Reference Figure 3 As shown in the figure, the conflict resolution of the knowledge graph in path planning is illustrated. The initial state represents two agents (A1, A2) and their respective goals, and the preliminary planned path shows potential collision points. During KG interaction, the knowledge graph (which may be a small fragment) displays the current position of the agents, the planned path, and the identified conflicts (e.g., "A1 is located at (x,y) at t+5", "A2 is located at (x,y) at t+5"). Then, conflict resolution / replanning is carried out. The decision module based on KG (which may show LLM interaction or GNN processing) identifies the conflicts and proposes revised paths for one or both agents (e.g., A1 waits, A2 chooses an alternative route). The final state represents the agents successfully reaching the goal without conflict.

[0057] S14, each task allocation knowledge fragment in the optimal task allocation scheme is sent to the corresponding intelligent agent for collaborative execution through a semantic communication protocol.

[0058] Furthermore, in step S14, sending each task allocation knowledge fragment in the optimal task allocation scheme to the corresponding intelligent agent for collaborative execution via a semantic communication protocol includes: S14-1, Based on the task-assigned knowledge fragment, query the shared knowledge graph to see if there is an intent conflict in the intent operation; S14-2, If an intent conflict is detected, arbitration is performed using a large language model combined with background knowledge in a shared knowledge graph, and a conflict resolution scheme is generated so that the conflicting agents can perform collaborative execution according to the conflict resolution scheme.

[0059] In this embodiment, structured semantic fragments (i.e., structured knowledge as the basic unit) are transmitted through a semantic communication protocol, rather than raw data / symbols (no longer text, numerical values, or simple instructions, etc.). That is, the knowledge graph subgraph query structure is transmitted, thereby providing technical support for establishing a unified cognitive view between agents. This allows agents to directly exchange semantically clear and reasonable knowledge segments, thereby significantly improving communication efficiency and accuracy.

[0060] By encapsulating knowledge graph subgraphs of entity, relation, and attribute triples, agents can directly exchange semantically clear and reasonable knowledge segments. For example, after completing a path planning task, agent A identifies "traffic congestion in region X." After processing by the experience-sharing graph center, it sends a knowledge graph subgraph composed of new reasoning conclusions "(region X, state, congestion)" and "(congestion, reasoning relation, avoidance path)" to other agents, thereby avoiding redundant computation and potential conflicts.

[0061] The semantic communication protocol uses the RDF / OWL standard semantic format for serialization (i.e., it uses the RDF / OWL standard semantic format to transmit the content of knowledge graph subgraphs), supplemented by lightweight compression strategies (such as knowledge subgraph summaries and node-selective transmission). The communication packet includes descriptive information (including trust tags, timestamps, and sender IDs) and the serialized subgraph to achieve contextual traceability of knowledge and source credibility control. For example, if agent B receives a subgraph segment containing a timestamp T and a credibility score of 0.85, it can determine whether this information overwrites existing conflict path information in the graph or whether a conflict resolution procedure should be triggered.

[0062] Furthermore, to improve communication efficiency and ensure semantic integrity, the system designs a multi-level semantic compression and transmission mechanism, combining the structural information of the knowledge graph with embedded semantic representations for encoding and transmission. Specifically, in bandwidth-constrained or fast-synchronization scenarios, the sender can transmit only the embedded representations of key entities in the knowledge graph subgraphs (such as Top-K node vectors selected based on graph attention mechanisms), combined with a small number of relation labels to form a compact semantic packet. The receiver utilizes the local knowledge graph and context descriptions to perform entity matching and concept localization based on embedding similarity, achieving semantic alignment. In scenarios with high semantic consistency requirements, the protocol supports the transmission of joint representations of structured triples and embedded representations. The embedded representation vectors are used to capture semantic proximity features, while the structured triples provide explicit reasoning paths. The receiver can achieve structural reconstruction and graph fusion of some knowledge fragments through embedding clustering, entity linking, KG completion modules, and context alignment strategies.

[0063] The system also introduces a knowledge-aware iterative retrieval mechanism, which dynamically generates query requests based on missing points in the receiver's knowledge graph during communication, avoiding redundant and invalid information, and optimizing contextual consistency through feedback loops. This mechanism can mitigate the risk of semantic drift in LLM inference and enhance the robustness of semantic alignment.

[0064] Reference Figure 4 The diagram shown is a structural schematic of a knowledge graph-based multi-agent collaboration enhancement device according to an embodiment of the present invention.

[0065] In this embodiment, the device 20 includes: The data acquisition unit 21 is used to acquire raw data from the external environment, preprocess and extract features from the raw data, generate knowledge triples and store them in the local knowledge graph, and submit the knowledge triples to the shared knowledge graph in an incremental manner based on the semantic communication protocol for knowledge updates. The task decomposition unit 22 is used to decompose the task to be executed into multiple sub-tasks by using a large language model and querying global knowledge in the shared knowledge graph and local knowledge in the local knowledge graph when a task to be executed is received. The scheme generation unit 23 is used to match multiple subtasks with the capabilities and availability of each agent using a bipartite graph minimum cost matching algorithm to generate a preliminary task allocation scheme, and to optimize the preliminary task allocation scheme using a large language model to generate an optimal task allocation scheme. The collaborative execution unit 24 is used to send each task allocation knowledge fragment in the optimal task allocation scheme to the corresponding intelligent agent for collaborative execution through a semantic communication protocol.

[0066] Each unit module of the device 20 can execute the corresponding steps in the above method embodiment, so the details of each unit module will not be elaborated here. Please refer to the description of the corresponding steps above for details.

[0067] This invention also provides a knowledge graph-based multi-agent collaboration enhancement device, which includes the knowledge graph-based multi-agent collaboration enhancement apparatus described above. The knowledge graph-based multi-agent collaboration enhancement apparatus can employ... Figure 4 The structure of the embodiment, correspondingly, can be executed Figure 1 The technical solutions of the method embodiments shown are similar in implementation principle and technical effect. For details, please refer to the relevant records in the above embodiments, which will not be repeated here.

[0068] The device includes: a mobile phone, digital camera, or tablet computer, or other device with a camera function; or a device with an image processing function; or a device with an image display function. The device may include components such as a memory, processor, input unit, display unit, and power supply.

[0069] The memory can be used to store software programs and modules. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, application programs required for at least one function (such as image playback function), etc.; the data storage area can store data created according to the use of the device. In addition, the memory can include high-speed random access memory, and can also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory can also include a memory controller to provide access to the memory for the processor and input units.

[0070] The input unit can be used to receive input numerical, character, or image information, and to generate keyboard, mouse, joystick, optical, or trackball signal inputs related to user settings and function control. Specifically, in addition to a camera, the input unit of this embodiment may also include a touch-sensitive surface (e.g., a touch screen) and other input devices.

[0071] The display unit can be used to display information input by the user or information provided to the user, as well as various graphical user interfaces of the device. These graphical user interfaces can be composed of graphics, text, icons, video, and any combination thereof. The display unit may include a display panel, optionally configured as an LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), or other similar display panel. Furthermore, a touch-sensitive surface may cover the display panel. When the touch-sensitive surface detects a touch operation on or near it, it transmits the information to the processor to determine the type of touch event. Subsequently, the processor provides corresponding visual output on the display panel based on the type of touch event.

[0072] This invention also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the memory described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into a device. The computer-readable storage medium stores at least one instruction, which is loaded and executed by a processor to implement... Figure 1 The illustrated method is a knowledge graph-based multi-agent collaboration enhancement method. The computer-readable storage medium can be a read-only memory, a hard disk, or an optical disk, etc.

[0073] This invention also provides a computer program product, including a computer program / instructions, which are loaded and executed by a processor to achieve [the desired result]. Figure 1 This paper presents a multi-agent collaboration enhancement method based on knowledge graphs.

[0074] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, device embodiments, and storage medium embodiments, since they are basically similar to the method embodiments, the descriptions are relatively simple, and relevant parts can be referred to the descriptions in the method embodiments.

[0075] Furthermore, in this document, 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 a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0076] The foregoing description illustrates and describes preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept by means of the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for enhancing multi-agent collaboration based on knowledge graphs, characterized in that, The method includes: The system acquires raw data from the external environment, preprocesses and extracts features from the raw data, generates knowledge triples which are stored in the local knowledge graph, and submits the knowledge triples to the shared knowledge graph in an incremental manner based on the semantic communication protocol for knowledge updates. When a task to be executed is received, the task is decomposed into multiple sub-tasks by using a large language model and querying global knowledge in the shared knowledge graph and local knowledge in the local knowledge graph. A bipartite graph minimum cost matching algorithm is used to match multiple subtasks with the capabilities and availability of each agent to generate a preliminary task allocation scheme. This preliminary task allocation scheme is then optimized using a large language model to generate an optimal task allocation scheme. The process of using the bipartite graph minimum cost matching algorithm to match multiple subtasks with the capabilities and availability of each agent to generate a preliminary task allocation scheme includes: Construct a bipartite graph with the set of subtasks and the set of agents as vertices; The matching cost weights of each edge in the bipartite graph are calculated based on the capability attributes, state attributes, and historical performance data of each agent recorded in the shared knowledge graph. The optimal subtask-agent allocation pair is solved by the minimum cost matching algorithm to generate the preliminary task allocation scheme; Each task allocation knowledge fragment in the optimal task allocation scheme is sent to the corresponding intelligent agent for collaborative execution via a semantic communication protocol.

2. The multi-agent collaboration enhancement method based on knowledge graphs according to claim 1, characterized in that, The step of submitting the knowledge triples to the shared knowledge graph incrementally for knowledge updating based on the semantic communication protocol includes: When semantic conflicts are detected between knowledge triples submitted by different agents, arbitration is performed by activating a conflict resolution agent based on a large language model, and the knowledge triples that meet the conditions are used as knowledge to update the shared knowledge graph after arbitration.

3. The multi-agent collaboration enhancement method based on knowledge graphs according to claim 2, characterized in that, The arbitration process, which involves activating a conflict resolution agent based on a large language model, includes: Send structured debate prompts to agents with semantic conflicts and collect evidence and reasoning paths provided by each agent; Arbitration of evidence and reasoning paths is based on large-scale language models by combining the credibility and confidence of information sources.

4. The method for enhancing multi-agent collaboration based on knowledge graphs according to claim 1, characterized in that, The step of optimizing the initial task allocation scheme using a large language model to generate the optimal task allocation scheme includes: The preliminary task allocation scheme, task context description, and agent metadata extracted from the shared knowledge graph are input into a large language model. The initial allocation scheme is evaluated and adjusted using a semantic bidding mechanism based on a contract network protocol using a large language model, generating interpretable reasons for task allocation and the optimal task allocation scheme.

5. The method for enhancing multi-agent collaboration based on knowledge graphs according to claim 1, characterized in that, The step of sending each task allocation knowledge fragment in the optimal task allocation scheme to the corresponding intelligent agent for collaborative execution via a semantic communication protocol includes: Based on the task-assigned knowledge fragment, query the shared knowledge graph to see if there is an intent conflict in the intent operation; If an intent conflict is detected, a large language model combined with background knowledge from a shared knowledge graph is used to arbitrate and generate a conflict resolution scheme, so that the conflicting agents can perform collaborative execution according to the conflict resolution scheme.

6. The method for enhancing multi-agent collaboration based on knowledge graphs according to claim 1, characterized in that, The semantic communication protocol supports multi-level transmission modes, including: In bandwidth-constrained scenarios, the sender transmits semantic packets consisting of embedded representations of key entities in the knowledge graph subgraph, and the receiver performs semantic reconstruction based on the local knowledge graph and context description. In scenarios with high semantic consistency, the sender transmits a semantic packet consisting of structured triples and embedded representations of knowledge graph subgraphs. The receiver then uses embedding clustering, entity linking, knowledge graph completion, and context alignment strategies to reconstruct the structure of knowledge fragments and fuse the graph.

7. A knowledge graph-based multi-agent collaboration enhancement device, characterized in that, The device includes: The data acquisition unit is used to acquire raw data from the external environment, preprocess and extract features from the raw data, generate knowledge triples and store them in the local knowledge graph, and submit the knowledge triples to the shared knowledge graph in an incremental manner based on the semantic communication protocol for knowledge updates. The task decomposition unit is configured to, upon receiving a task to be executed, decompose the task into multiple sub-tasks using a large language model and by querying global knowledge in a shared knowledge graph and local knowledge in a local knowledge graph; wherein, the task decomposition unit is further configured to: Construct a bipartite graph with the set of subtasks and the set of agents as vertices; The matching cost weights of each edge in the bipartite graph are calculated based on the capability attributes, state attributes, and historical performance data of each agent recorded in the shared knowledge graph. The optimal subtask-agent allocation pair is solved by the minimum cost matching algorithm to generate the preliminary task allocation scheme; The scheme generation unit is used to match multiple subtasks with the capabilities and availability of each agent using a bipartite graph minimum cost matching algorithm to generate a preliminary task allocation scheme, and to optimize the preliminary task allocation scheme using a large language model to generate an optimal task allocation scheme. The collaborative execution unit is used to send each task allocation knowledge fragment in the optimal task allocation scheme to the corresponding intelligent agent for collaborative execution through a semantic communication protocol.

8. A multi-agent system, characterized in that, The system includes: The perception module is used to acquire raw data from the external environment and send it to the local knowledge graph module; The local knowledge graph module is used to preprocess and extract features from raw data, generate knowledge triples as local knowledge of the agent and store them, and perform knowledge reasoning and context understanding according to the needs of the decision module. The shared knowledge graph module is used to receive knowledge update requests from various agents, realize knowledge updates by incrementally constructing knowledge triples, and receive knowledge query requests from various agents to provide global knowledge. The decision-making module is used to obtain local knowledge from the local knowledge graph module and global knowledge from the shared knowledge graph module to perform task decomposition and generate execution instructions for task allocation schemes. The communication module is used to encapsulate execution instructions into corresponding knowledge fragments and send them to the corresponding intelligent agent through a semantic communication protocol, as well as to submit knowledge triples to the shared knowledge graph.

9. A knowledge graph-based multi-agent collaboration enhancement device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory, the computer program being executed by the processor to implement the steps of a knowledge graph-based multi-agent collaboration enhancement method as described in any one of claims 1 to 6.

Images