CN-122021708-A - Multi-agent cooperative calling method and system based on knowledge-graph correlation

CN122021708ACN 122021708 ACN122021708 ACN 122021708ACN-122021708-A

Abstract

The invention discloses a multi-agent cooperative calling method and a system based on knowledge graph association, which relate to the technical field of data processing, and the method comprises the steps of carrying out semantic analysis on a task original demand text to extract a task feature vector; the method comprises the steps of firstly, obtaining a task feature vector, then obtaining a semantic knowledge graph of a plurality of agents, then calling the semantic knowledge graph of the plurality of agents to match, identifying a target agent set and constructing a local calling sub-graph, then dynamically correcting initial association weights of the sub-graph based on the task feature vector to obtain a real-time association matrix, then quantifying collaboration priority and dependency strength, executing redundancy elimination based on semantic projection gradient, generating a calling sequence by combining a bottom layer physical load state, finally obtaining real-time physical load and resource occupation conditions during execution, executing conflict scheduling based on a dynamic access quota model, and issuing an adjustment instruction. The invention has the advantages of accurate coordination, dynamic self-adaption and good calling effect.

Inventors

HU LEI
HE ZHAOFENG

Assignees

四川省人工智能研究院
成都博诚智医软件技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. A multi-agent cooperative calling method based on knowledge-graph association is characterized by comprising the following steps: acquiring an original demand text of a task to be executed, and executing semantic analysis on the original demand text to extract a task feature vector; A preset multi-agent semantic knowledge graph is called, the task feature vector is matched with each agent node in the multi-agent semantic knowledge graph, and a target agent set is identified; constructing a local calling sub-graph according to preset association edges among all agent nodes in the target agent set, and dynamically correcting initial association weights in the local calling sub-graph based on the task feature vector to acquire a real-time association matrix; quantifying the collaboration priority and the dependency strength among the intelligent agent nodes according to the real-time correlation matrix, performing redundancy elimination based on semantic projection gradient, and generating a multi-intelligent agent collaboration calling sequence by combining a bottom layer physical load state and the dependency strength; and acquiring the real-time physical load state and the logic resource occupation condition of each agent node when executing the multi-agent cooperative call sequence, executing conflict scheduling based on a dynamic access quota model and issuing a cooperative adjustment instruction.
2. The knowledge-graph-association-based multi-agent collaboration calling method of claim 1, wherein the obtaining the original demand text of the task to be performed and performing semantic parsing on the original demand text to extract the task feature vector comprises: Identifying task verbs, operation objects and constraint conditions in the original demand text, and mapping the task verbs, the operation objects and the constraint conditions to a high-dimensional semantic space respectively to obtain corresponding verb component vectors, object component vectors and constraint component vectors; Acquiring a verb weight coefficient, an object weight coefficient and a constraint weight coefficient which are preset based on semantic importance and the sum of which is one; And performing weighted fusion operation on the verb component vector, the object component vector and the constraint component vector by using the verb weight coefficient, the object weight coefficient and the constraint weight coefficient to acquire the task feature vector representing the global semantic feature of the task to be executed.
3. The knowledge-graph-association-based multi-agent cooperative call method according to claim 1, wherein the dynamically modifying the initial association weight in the local call sub-graph based on the task feature vector to obtain a real-time association matrix comprises: calculating cosine alignment degree of a sum vector of the attribute feature vectors of the task feature vector and two connected agent nodes in the local call subgraph, and taking the cosine alignment degree as semantic projection intensity of the task feature vector between the two connected agent nodes; constructing a forward gain proportion by using the product of the semantic projection intensity and a preset adjustment coefficient; Multiplying the initial association weight between the two agent nodes by the sum of a forward gain ratio to obtain corrected real-time association weights, and constructing the real-time association matrix by arranging all corrected real-time association weights according to node indexes.
4. The knowledge-graph-correlation-based multi-agent co-modulation method of claim 1, wherein the performing redundancy culling based on semantic projection gradients comprises: Calculating cosine similarity of attribute feature vectors of any two agent nodes in the target agent set; calculating absolute values of differences between projection scalar quantities of attribute feature vectors of the two agent nodes on the task feature vectors respectively, and taking the absolute values as projection gradient differences; and when the cosine similarity is larger than a preset redundancy threshold and the projection gradient difference is smaller than a preset gradient interference threshold, judging that functional redundancy exists between the two agent nodes, reserving the agent nodes with higher corresponding corrected real-time association weights in the real-time association matrix, and executing redundancy elimination operation.
5. The knowledge-graph-correlation-based multi-agent co-modulation method of claim 1, wherein the generating a multi-agent co-modulation sequence in combination with the underlying physical load state and the dependency strength comprises: Acquiring the release time of a data interaction handle of an upstream intelligent agent node and the real-time physical comprehensive load rate of a downstream intelligent agent node; Multiplying a preset reference system response constant by an exponential term taking a natural constant as a base and taking the product of a preset hardware penalty coefficient and the real-time physical comprehensive load rate as an index to obtain load penalty response time; Dividing the load penalty response time by the corrected real-time association weight between the upstream agent node and the downstream agent node to obtain additional delay time; and adding the additional delay time to the data interaction handle release time, determining the data interaction handle release time as a starting time stamp of the downstream agent node, and generating the multi-agent cooperative call sequence according to the time sequence arrangement of the starting time stamp.
6. The knowledge-graph-association-based multi-agent co-modulation method of claim 1, wherein the performing the conflict scheduling and issuing co-modulation instructions based on the dynamic access quota model comprises: when a plurality of agent nodes are predicted to compete for the same logic resource within the same preset time period, the corresponding corrected real-time association weight in the real-time association matrix is called as a heuristic factor to calculate the competition priority; Adding one to the corrected real-time association weight to obtain natural logarithms to obtain logarithmic level increments; Multiplying a preset system basic access quota issuing rate by a sum of a logarithmic increment and the sum of the logarithmic increment, and calculating to obtain a dynamic quota supplementing rate exclusive to each competitive intelligent agent node; And opening the bottom physical communication channel at a corresponding exclusive dynamic quota supplement rate, and distributing logic resource access permission to the intelligent agent node according to the competition priority.
7. A multi-agent co-modulation system based on knowledge-graph correlation, the system comprising: the task analysis module is used for acquiring an original demand text of a task to be executed, and executing semantic analysis on the original demand text to extract a task feature vector; The map management module is used for retrieving a preset multi-agent semantic knowledge map, wherein the multi-agent semantic knowledge map comprises a plurality of agent nodes, functional attribute labels corresponding to the agent nodes and preset associated edges among the nodes; The association quantization module is used for matching the task feature vector with each agent node in the multi-agent semantic knowledge graph, identifying a target agent set, constructing a local calling subgraph according to the target agent set, dynamically correcting the initial association weight based on the task feature vector, and acquiring a real-time association matrix; the sequence generation module is used for quantifying the collaboration priority and the dependence intensity among the agent nodes according to the real-time correlation matrix, executing redundancy elimination based on semantic projection gradient, and generating a multi-agent collaboration calling sequence by combining a bottom layer physical load state and the dependence intensity; And the conflict scheduling module is used for acquiring the real-time physical load state and the logic resource occupation condition of each agent node when the multi-agent cooperative call sequence is executed, executing conflict scheduling based on the dynamic access quota model and issuing a cooperative adjustment instruction.
8. The knowledge-graph correlation-based multi-agent co-modulation system of claim 7, wherein the task parsing module is further configured to: Identifying task verbs, operation objects and constraint conditions in the original demand text, and mapping the task verbs, the operation objects and the constraint conditions to a high-dimensional semantic space respectively to obtain corresponding verb component vectors, object component vectors and constraint component vectors; Acquiring a verb weight coefficient, an object weight coefficient and a constraint weight coefficient which are preset based on semantic importance and the sum of which is one; And performing weighted fusion operation on the verb component vector, the object component vector and the constraint component vector by using the verb weight coefficient, the object weight coefficient and the constraint weight coefficient to acquire the task feature vector representing the global semantic feature of the task to be executed.
9. The knowledge-graph correlation-based multi-agent co-modulation system of claim 7, wherein the correlation quantization module is further configured to: calculating cosine alignment degree of a sum vector of the attribute feature vectors of the task feature vector and two connected agent nodes in the local call subgraph, and taking the cosine alignment degree as semantic projection intensity of the task feature vector between the two connected agent nodes; constructing a forward gain proportion by using the product of the semantic projection intensity and a preset adjustment coefficient; Multiplying the initial association weight between the two agent nodes by the sum of a forward gain ratio to obtain corrected real-time association weights, and constructing the real-time association matrix by arranging all corrected real-time association weights according to node indexes.
10. The knowledge-graph correlation-based multi-agent co-modulation system of claim 7, wherein the sequence generation module is further configured to: Calculating cosine similarity of attribute feature vectors of any two agent nodes in the target agent set; calculating absolute values of differences between projection scalar quantities of attribute feature vectors of the two agent nodes on the task feature vectors respectively, and taking the absolute values as projection gradient differences; and when the cosine similarity is larger than a preset redundancy threshold and the projection gradient difference is smaller than a preset gradient interference threshold, judging that functional redundancy exists between the two agent nodes, reserving the agent nodes with higher corresponding corrected real-time association weights in the real-time association matrix, and executing redundancy elimination operation.

Description

Multi-agent cooperative calling method and system based on knowledge-graph correlation Technical Field The invention relates to the technical field of data processing, in particular to a multi-agent cooperative calling method and system based on knowledge-graph correlation. Background In the field of complex industrial automation and distributed computing, a multi-agent collaborative call system is a core infrastructure for realizing high-dimensional business process automation. However, the existing multi-agent cooperative technology mainly relies on shallow keyword matching when facing unstructured natural language instructions, and lacks deep semantic deconstructing and multidimensional vector fusion capabilities for task actions, objects and complex constraint conditions, so that initial task distribution and node matching accuracy are insufficient. In the aspect of collaborative topology construction, a traditional knowledge graph generally adopts static association weight, projection intensity cannot be dynamically calculated according to the semantic context of a current specific task, and association gain among nodes can be adaptively adjusted, so that a calling link is stiff and has a slow response. Meanwhile, in a huge intelligent agent cluster, the existing redundancy elimination mechanism is limited to single node attribute similarity comparison, and complementary intelligent agents with different actual force applying directions are easily misjudged as redundant nodes and eliminated under a complex task background, so that serious function miskilling and calculation force waste are caused. More importantly, the existing collaborative scheduling sequence generation and queuing mechanism is generally limited to logic deduction of a pure software layer, and the causal mapping relation between the upper semantic priority and the real-time physical state of the bottom hardware is completely split. When the bottom physical node is close to full load, the system still forcibly distributes tasks according to a static logic time sequence, and when a plurality of agents compete for the same common logic resource, a dynamic bottom access quota management and control mechanism based on core service weight is lacking, which easily causes serious hardware response exponential delay, execution deadlock and data throughput collapse of a core service link, and finally leads the whole multi-agent cooperative architecture to be difficult to ensure physical feasibility and overall stability under a dynamic high concurrency scene. Disclosure of Invention Aiming at the technical problems recorded in the background technology, the invention provides a multi-agent cooperative calling method and system based on knowledge-graph correlation. A multi-agent cooperative calling method based on knowledge graph association comprises the steps of obtaining an original demand text of a task to be executed, executing semantic analysis on the original demand text to extract a task feature vector, calling a preset multi-agent semantic knowledge graph, matching the task feature vector with each agent node in the multi-agent semantic knowledge graph, identifying a target agent set, constructing a local calling sub-graph according to preset association edges among the agent nodes in the target agent set, dynamically correcting initial association weights in the local calling sub-graph based on the task feature vector to obtain a real-time association matrix, quantifying cooperative priority and dependency strength among the agent nodes according to the real-time association matrix, executing redundancy elimination based on a semantic projection gradient, generating a multi-agent cooperative calling sequence based on a bottom physical load state and the dependency strength, obtaining real-time physical load states and logic resource occupation conditions of the agent nodes when executing the multi-agent cooperative calling sequence, executing conflict scheduling based on a dynamic quota model, and issuing a cooperative adjustment instruction. The method comprises the steps of identifying a task verb, an operation object and constraint conditions in an original demand text, mapping the task verb, the operation object and the constraint conditions to a high-dimensional semantic space respectively, obtaining a corresponding verb component vector, an object component vector and a constraint component vector, obtaining a verb weight coefficient, an object weight coefficient and a constraint weight coefficient which are preset based on semantic importance and are one in sum, and carrying out weighted fusion operation on the verb component vector, the object component vector and the constraint component vector by utilizing the verb weight coefficient, the object weight coefficient and the constraint weight coefficient to obtain the task feature vector representing global semantic features of the task to be executed. Optionally, dynamically c