Heterogeneous large language model multi-agent-oriented multi-scale task collaborative arrangement method and equipment

Abstract

The invention discloses a heterogeneous large language model multi-agent-oriented multi-scale task collaborative orchestration method and device, and the method comprises the steps: firstly obtaining a to-be-orchestrated task processing demand and the task adaptation capability of each agent, and constructing a heterogeneous interaction graph containing agent nodes and task nodes; thirdly, clustering and grouping nodes in the graph according to the task adaptation capability of the intelligent agent and the similarity degree of task processing requirements; then, performing coarse-grained processing on the grouped graphs by using a pre-trained multi-scale collaborative task arrangement model to obtain a coarse-grained collaborative arrangement scheme; and inputting the task nodes into a task scheduling priority model to obtain priorities, and carrying out fine-grained processing on the coarse-grained scheme by using the priorities to obtain a fine-grained collaborative arrangement scheme between the intelligent agent and the corresponding task. The invention aims to solve the problem of low scheduling efficiency caused by agent isomerism, high communication overhead and task dynamic complexity in task arrangement.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent agents, and particularly relates to a multi-scale task coordination arrangement method and device for heterogeneous large language model multi-agents. BACKGROUND

[0002] The agent based on the large language model (large language model agent) has gradually become a new focus in the field of multi-agent system research. The large language model agent has strong reasoning ability, autonomous learning ability and processing ability for complex tasks, and can realize higher level of autonomous optimization and decision-making in a high-dimensional dynamic environment, thereby improving the overall intelligence level and adaptability of the multi-agent system. However, in actual application scenarios, the multi-agent system often faces strong system heterogeneity, and the agents have differences in computing power, communication performance, cognitive behavior, etc. The communication overhead is large, and the information interaction between the agents needs to consume a large amount of resources. The task is dynamic and complex, and the demand and state of the task change over time. These problems make the task arrangement work extremely complex, not only the computing and communication capabilities of the agents need to be fully considered, but also the dynamic changes of the environment and the quality of service demand of the task need to be closely adapted. Especially in an open system, as the number of agents and the size of the task continue to grow, the traditional task arrangement strategy is difficult to balance between computing efficiency, system scalability and real-time responsiveness. Therefore, to fully exert the cooperative advantages of the large language model agent, an efficient task arrangement and scheduling strategy has become a core problem to be solved, and how to construct an efficient, flexible and scalable task arrangement mechanism in a large-scale, heterogeneous and variable environment has become a key challenge in the development process of the multi-agent system.

[0003] To address the above problems, existing research has proposed various task scheduling and dispatching strategies to improve the overall performance of multi-agent systems. Current common methods include different approaches such as reinforcement learning, hierarchical scheduling models, mathematical modeling, and heuristic optimization. For example, Beijing University of Posts and Telecommunications proposed a multi-agent near-optimal policy optimization-based collaborative task scheduling method, which uses the near-optimal policy optimization mechanism in reinforcement learning to improve the efficiency of task allocation for agents. Hunan University proposed a hierarchical scheduling model for multi-task multi-agent allocation problems, which divides the agent execution unit through a hierarchical structure to effectively improve the manageability and scheduling flexibility of the system. Massachusetts Institute of Technology constructed a multi-agent autonomous model based on opinion dynamics and multi-objective behavior optimization, proposed a theoretical analysis framework for group behavior coordination, and enhanced the autonomous decision-making ability and task scheduling efficiency of multi-agent systems in dynamic environments. The Institute of Automation of the Chinese Academy of Sciences proposed an agent coordination strategy based on graph generation, which uses a graph generator and a graph-driven distributed decision-making mechanism to achieve efficient management and scheduling of agent resources in complex dynamic environments. The research team at Tsinghua University proposed a dynamic parameter sharing method that uses a self-supervised learning mechanism to extract implicit behavior features from agent action trajectories, optimizing the efficiency of task scheduling strategies in multi-agent systems.

[0004] Although existing research has made some achievements, when facing real large language model multi-agent systems, the following problems still exist. On the one hand, existing methods generally fail to fully consider the heterogeneous characteristics of agents in terms of computing power, communication performance, and cognitive behavior, making it difficult to effectively meet the scheduling needs in complex dynamic environments. On the other hand, as the system size continues to expand, most methods expose defects such as high computational complexity, poor real-time response, and weak scalability, making it difficult to guarantee scheduling performance in large-scale deployment. At the same time, some methods that rely on mathematical modeling and heuristic rules cannot effectively extract the deep dynamic interaction between large language model agents and tasks, resulting in limited effectiveness of scheduling strategies in actual scenarios. Therefore, constructing a task and agent matching method that can balance heterogeneity and large-scale system scalability is still a key technical problem that needs to be broken through in the current research field. SUMMARY

[0005] In view of the problems in the prior art, the present application provides a multi-scale task collaborative scheduling method and equipment for heterogeneous large language model multi-agent, which aims to solve the problem of low scheduling efficiency caused by agent heterogeneity, large communication overhead, and task dynamic complexity in task scheduling.

[0006] To solve the above technical problems, the present application is implemented by the following technical solutions:

[0007] According to a first aspect of the present application, a multi-scale task collaborative scheduling method for heterogeneous large language model multi-agent is provided, comprising:

[0008] Obtain the processing requirements of the tasks to be orchestrated and the task adaptation capabilities of each agent;

[0009] Based on the processing requirements of the tasks to be orchestrated and the task adaptation capabilities of each agent, a heterogeneous interaction graph between agents and tasks to be orchestrated is constructed. The heterogeneous interaction graph includes agent nodes and task nodes.

[0010] Based on the similarity of the agents' task adaptation capabilities and the similarity of the processing requirements of the tasks to be orchestrated, the agent nodes and task nodes in the heterogeneous interaction graph are clustered and grouped to obtain the heterogeneous interaction graph between the agents and the tasks to be orchestrated after clustering.

[0011] A coarse-grained multi-scale collaborative task orchestration scheme is obtained by using a pre-trained multi-scale collaborative task orchestration model to perform coarse-grained processing on the heterogeneous interaction graph between the clustered agents and the tasks to be orchestrated. The coarse-grained multi-scale collaborative task orchestration scheme is a collaborative orchestration scheme between the clustered agents and the tasks to be orchestrated.

[0012] The task nodes are input into a pre-built task scheduling priority model to obtain task priorities. The task priorities are then used to refine the coarse-grained multi-scale collaborative task orchestration scheme to obtain a fine-grained multi-scale collaborative task orchestration scheme. The fine-grained multi-scale collaborative task orchestration scheme is a collaborative orchestration scheme between the agent and the corresponding task to be orchestrated.

[0013] In one possible implementation of the first aspect, the processing requirements of the task to be orchestrated include computing resource requirements, storage requirements, and urgency requirements;

[0014] The task adaptation capability of the intelligent agent includes the core capability of the intelligent agent to complete a specific task, the currently available computing and communication resources, and the current coordinate information in physical or logical space.

[0015] In one possible implementation of the first aspect, the construction of a heterogeneous interaction graph between the agent and the task to be orchestrated, based on the processing requirements of the task to be orchestrated and the task adaptation capability of each agent, specifically involves:

[0016] According to the processing requirements of the tasks to be orchestrated, the tasks to be orchestrated are decomposed into multiple subtasks, each subtask is abstracted into a task node, and each task node is assigned corresponding attributes according to the processing requirements of each subtask.

[0017] Each agent is abstracted into an agent node, and corresponding attributes are assigned to the agent node based on the task adaptation capability of each agent.

[0018] According to the correlation degree between the sub-tasks, edges are constructed between the task nodes, and the edges between the task nodes are assigned with corresponding attributes according to the execution order of the task nodes, and the correlation degree between the sub-tasks is determined according to the dependency degree and the priority between the sub-tasks;

[0019] According to the correlation degree between the agents, edges are constructed between the agent nodes, and the edges between the agent nodes are assigned with corresponding attributes according to the coordination efficiency of the agents, and the correlation degree between the agents is determined according to the communication rate and the related degree of processing tasks between the agents;

[0020] According to the adaptation degree between the agents and the sub-tasks, edges are constructed between the task nodes and the agent nodes, and the edges between the task nodes and the agent nodes are assigned with corresponding attributes.

[0021] In a possible implementation manner of the first aspect, the agent nodes and the task nodes in the heterogeneous interaction graph are clustered and grouped according to the similarity degree of the task adaptation capability of the agents and the similarity degree of the processing requirement of the to-be-scheduled tasks respectively, to obtain a clustered and grouped heterogeneous interaction graph between the agents and the to-be-scheduled tasks, and the clustering and grouping specifically includes:

[0022] 1) extracting attributes of each agent node, abstracting the attributes into an agent feature vector, and performing standardization processing on the agent feature vector;

[0023] 2) based on the agent feature vector after the standardization processing, a similarity matrix between the agent nodes is constructed by using cosine similarity measurement;

[0024] 3) a label propagation clustering method is used to construct a weighted edge graph based on the original heterogeneous interaction graph by using the similarity matrix between the agent nodes;

[0025] 4) each agent node in the weighted edge graph is assigned with a unique and mutually exclusive initial label, and each agent node is independently grouped;

[0026] 5) the agent nodes in the weighted edge graph are traversed in a random order, and the label of each agent node is updated to the highest-vote label by using neighbor labels and edge weights and using a weighted voting method;

[0027] 6) step 5) is repeated iteratively until the maximum iteration number is reached and the label is completely stable, and finally the agent nodes with the same label are aggregated into a group;

[0028] 7) extracting attributes of each task node, abstracting the attributes into a task feature vector, and performing standardization processing on the task feature vector;

[0029] 8) Based on the standardized task feature vector, the cosine similarity measure is used to construct the similarity matrix between task nodes;

[0030] 9) The label propagation clustering method is used to construct a weighted edge graph based on the original heterogeneous interaction graph using the similarity matrix between task nodes;

[0031] 10) Each task node in the weighted edge graph is assigned a unique and mutually exclusive initial label, and each task node is independently grouped;

[0032] 11) The task nodes in the weighted edge graph are traversed in random order, and the neighbor label and edge weight are used to update the label of each task node to the highest vote label using the weighted voting method;

[0033] 12) Repeat step 11) until the maximum number of iterations is reached and the label is completely stable, and finally aggregate the task nodes with the same label into a group;

[0034] 13) Abstract the agent nodes and task nodes belonging to the same group into super nodes, and reconstruct the structure based on the original heterogeneous interaction graph to obtain the heterogeneous interaction graph between the clustered agent and the to-be-scheduled task.

[0035] In a possible implementation of the first aspect, the cosine similarity measure is used to construct the similarity matrix between the agent nodes, specifically:

[0036] Based on the agent matrix Calculate the similarity S between the agent nodes ij = v i · v j / ‖v i ‖·||v j ||;

[0037] Based on the similarity between the agent nodes, the similarity matrix between the agent nodes is constructed;

[0038] wherein represents the feature representation of the nth agent, with a dimension of h; S ij represents the similarity between agent node i and agent node j; v i represents the feature vector of agent node i, with a dimension of h; v j represents the feature vector of agent node j, with a dimension of h; ‖v i ‖ represents the norm of v i ; ||v j || represents the norm of v j ;

[0039] The cosine similarity measure is used to construct a similarity matrix between task nodes, specifically:

[0040] Based on the task matrix Calculate the similarity G between task nodes ij = u i · u j / ‖u i ‖·||u j ||;

[0041] Based on the similarity between the task nodes, a similarity matrix between the task nodes is constructed.

[0042] In the formula, indicates the feature representation of the nth task, with a dimension of h; G ij indicates the similarity between task node i and task node j; u i indicates the feature vector of task node i, with a dimension of h; u j indicates the feature vector of task node j, with a dimension of h; ‖u i ‖ indicates the norm of u i ; ||u j || indicates the norm of u j .

[0043] In a possible implementation of the first aspect, the weighted voting method is used to update the label of each agent node to the highest vote label, specifically:

[0044]

[0045] In the formula, Score(c) represents the number of votes for the current label of each agent node; W ij represents the edge weight between the current agent node i and agent node j; j∈N(i) represents only traversing the agent nodes directly connected to agent node i; L j =c represents only counting the neighbor agent nodes j currently holding label c; the new label of agent node i is updated to

[0046] The weighted voting method is used to update the label of each task node to the highest vote label, specifically:

[0047]

[0048] In the formula, Score(z) represents the number of votes for the current label of each task node; Q ij represents the edge weight between the current task node i and task node j; j∈M(i) represents only traversing the task nodes directly connected to task node i; L j=z represents only counting the neighbor task nodes j holding the label z at present; the new label of the task node i is updated as

[0049] In a possible implementation manner of the first aspect, the training process of the multi-scale collaborative scheduling model comprises:

[0050] The graph neural network is trained using the agent-task cooperation dataset AirSim to obtain a multi-scale collaborative scheduling model capable of outputting an agent-task matching relationship; a loss function used in the graph neural network training process comprehensively considers task processing accuracy, task completion time and agent processing energy consumption, and the loss function is specifically:

[0051]

[0052] In the formula, k represents the kth task scheduling scheme; represents the time loss of the agent system executing the task; E ed represents the energy consumption of the agent e processing the task d; represents the accuracy loss of the agent system executing the task; represents the agent task allocation state, The value of is 1 or 0, which is output in the multi-scale collaborative task scheduling scheme, 1 represents that the task allocation is successful, and 0 represents that the task is not allocated; α represents the weight of the accuracy loss; β represents the weight of the time loss; γ represents the weight of the energy consumption loss.

[0053] In a possible implementation manner of the first aspect, the task scheduling priority model comprises an optimization target and a constraint condition:

[0054] The optimization target is:

[0055]

[0056] The constraint condition is:

[0057] In the formula, x d represents the start time of the task d; y ed represents whether the task d is placed on the agent e; W d represents the priority of the task d; C d represents the completion time of the task d; T ed represents the execution time of the task d on the agent e; y ef represents whether the task f is placed on the agent e; T ef represents the execution time of the task f on the agent e; f = argmin f＝d,k {x f represents determining the task with an earlier start time from the task d and the task k; yed = y ek = 1 indicates that the task d and the task k are assigned to the same agent e.

[0058] In a possible implementation manner of the first aspect, the coarse-grained multi-scale collaborative task scheduling scheme is processed in a fine-grained manner by using the task priority to obtain a fine-grained multi-scale collaborative task scheduling scheme, and the processing specifically includes:

[0059] Based on the matching relationship between the agent groups and the task groups determined in the coarse-grained multi-scale collaborative task scheduling scheme, the subtasks in the task groups are mapped to the agents in the agent groups, and a fine-grained allocation problem between the task subgraph and the agent subgraph is constructed;

[0060] Based on the task priority and in combination with the attributes of the agent nodes, the subtasks in each task group are sorted;

[0061] On the basis of the sorting result, the specific subtasks are assigned to the optimal agents in the agent groups, while the currently available computing and communication resource capabilities and the current coordinate information in the physical or logical space are considered;

[0062] After completing all the fine-grained mapping, a final fine-grained multi-scale collaborative task scheduling scheme is output.

[0063] According to the second aspect of the present application, a computer device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the multi-scale task collaborative scheduling method for heterogeneous large language model multi-agent when executing the computer program.

[0064] Compared with the prior art, the present application has at least the following beneficial effects:

[0065] The application provides a multi-scale task collaborative arrangement method for a heterogeneous large language model multi-agent, which comprehensively considers the task adaptation capability of the agent and the processing requirement of the to-be-arranged task by constructing a multi-scale task arrangement mechanism based on a heterogeneous interaction graph, constructs a heterogeneous interaction graph by using the agent node and the task node, fully models the heterogeneity between individuals in the multi-agent system in terms of calculation, communication and cognitive ability, enables the task arrangement to more accurately match the agent and the task, and allocates the most suitable task according to the characteristics of different agents, thereby improving the accuracy of task scheduling and better adapting to complex and variable actual scenes. In combination with a clustering mechanism and a multi-scale collaborative model, the nodes in the heterogeneous interaction graph are clustered and grouped according to the similarity of the agent task adaptation capability and the similarity of the to-be-arranged task processing requirement, and the multi-scale collaborative task arrangement model pre-trained is used to perform coarse-grained processing on the clustered and grouped heterogeneous interaction graph. This processing mode decomposes the large-scale scheduling problem into multiple small-scale sub-problems, effectively reduces the computational complexity, reduces the system burden, and enables the system to run more efficiently when processing large-scale tasks and agents. The application adopts a coarse-fine-grained phased scheduling mechanism, obtains the clustered and grouped collaborative arrangement scheme through coarse-grained processing, uses a task scheduling priority model to perform priority sorting on the task node, and then performs fine-grained processing on the coarse-grained scheme to obtain the collaborative arrangement scheme between the agent and the corresponding to-be-arranged task. This phased scheduling mechanism enables the system to flexibly adjust the task arrangement strategy according to the real-time priority and dynamic change of the task, not only improves the adaptability of the system in a complex dynamic environment, but also ensures that the system can timely respond to the change of the task requirement, and enhances the expansion and real-time response capability of the scheduling strategy.

[0066] In order to make the above objectives, characteristics and advantages of the present application more apparent, clear and easy to understand, a preferred embodiment is described below in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to more clearly illustrate the technical solutions in the specific embodiments of the present application, the drawings needed in the description of the specific embodiments will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0068] Figure 1 A flow chart of a multi-scale task collaborative arrangement method for a heterogeneous large language model multi-agent according to an embodiment of the present application;

[0069] Figure 2 A flow and structure representation of abstracting an agent and a task into a node graph;

[0070] Figure 3 a coarse-grained multi-scale collaborative task scheduling scheme, a fine-grained multi-scale collaborative scheduling scheme construction process, and a structure representation;

[0071] Figure 4 for internal training and operation processes of a graph neural network. DETAILED DESCRIPTION

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described below in detail with reference to the drawings. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0073] The method is directed to a heterogeneous large language model multi-agent system, and solves the problem of low scheduling efficiency caused by agent heterogeneity (differences in computing power, communication performance, and cognitive behavior), large communication overhead, and dynamic complexity of tasks in task scheduling. The system includes multiple large language model agents with autonomous decision-making capabilities, which need to collaboratively process tasks that change over time in a high-dimensional dynamic environment while meeting quality of service requirements.

[0074] In combination with Figures 1 to 4 The embodiments of the present application provide a multi-scale task collaborative scheduling method for a heterogeneous large language model multi-agent, which specifically includes the following steps:

[0075] Step 1, obtaining the processing requirements of the tasks to be scheduled and the task adaptation capabilities of each agent.

[0076] Specifically, the processing requirements of the tasks to be scheduled include computing resource requirements (such as CPU / GPU computing power requirements), storage requirements (such as memory / disk space), and urgency requirements (such as deadline constraints).

[0077] The task adaptation capabilities of the agent include the core capabilities of the agent to complete specific tasks (such as natural language processing, image recognition, and other skills), the current available computing and communication resource capabilities (such as current available computing resources, communication bandwidth), and the current coordinate information in the physical or logical space (physical location or logical topology coordinates).

[0078] For example, in an industrial production scenario, there are multiple heterogeneous large language model agents, including agent A responsible for quality detection, agent B responsible for equipment control, agent C responsible for logistics scheduling, etc. At the same time, there is a series of tasks to be scheduled, such as product detection tasks, equipment parameter adjustment tasks, material handling tasks, etc.

[0079] For the processing requirements of the to-be-scheduled task, taking a product detection task as an example, the computing resource requirement thereof is to call a specific image recognition algorithm and occupy a certain GPU computing resource; the storage requirement thereof is to store image data in a detection process and occupy a certain storage space; and the urgency requirement thereof is to complete detection within 10 minutes after a product is produced.

[0080] For the task adaptation capability of the agent, the core capability of the agent A to complete a specific task is to have an image recognition capability, the currently available computing and communication resource capability is that the GPU resource is 50% remaining, the network bandwidth is 100 Mbps, and the current coordinate information in the physical space is that the agent is located in a detection area of a production workshop. The task adaptation capabilities of the agent B and the agent C are determined in a similar manner. Through a system interface or a data acquisition module, the processing requirements of the to-be-scheduled task and the information about the task adaptation capability of each agent are obtained.

[0081] Step 2. According to the processing requirements of the to-be-scheduled task and the task adaptation capability of each agent, a heterogeneous interaction graph between the agent and the to-be-scheduled task is constructed, and the heterogeneous interaction graph comprises an agent node and a task node.

[0082] In an implementable manner, according to the processing requirements of the to-be-scheduled task and the task adaptation capability of each agent, the heterogeneous interaction graph between the agent and the to-be-scheduled task is constructed, specifically comprising:

[0083] Step 2.1. According to the processing requirements of the to-be-scheduled task, the to-be-scheduled task is decomposed into a plurality of subtasks, each subtask is abstracted into a task node, and each task node is given a corresponding attribute according to the processing requirement of each subtask.

[0084] For example, each task node is given a corresponding attribute {computing requirement, storage requirement, urgency}.

[0085] Illustratively, according to the processing requirements of the to-be-scheduled task, the product detection task is decomposed into a plurality of subtasks such as image acquisition, image preprocessing, image recognition, etc. Each subtask is abstracted into a task node, such as an image acquisition task node, an image preprocessing task node, an image recognition task node, etc. And according to the processing requirement of each subtask, each task node is given a corresponding attribute, such as the attributes of the image acquisition task node including the type of acquisition device, the acquisition frequency, etc.; the attributes of the image preprocessing task node including the type of preprocessing algorithm, the processing time, etc.; and the attributes of the image recognition task node including the type of recognition model, the accuracy requirement, etc.

[0086] Step 2.2. Each agent is abstracted into an agent node, and a corresponding agent node is given a corresponding attribute according to the task adaptation capability of each agent.

[0087] For example, the corresponding agent node is endowed with the corresponding attribute {core ability, available resource, spatial coordinate}.

[0088] For example, each agent is abstracted into an agent node, such as agent A node, agent B node, agent C node, etc. According to the task adaptation ability of each agent, the corresponding agent node is endowed with the corresponding attribute, such as the attribute of agent A node including image recognition ability level, GPU resource remaining amount, network bandwidth, etc.; agent B node and agent C node are also endowed with similar attributes.

[0089] Step 2.3, according to the association degree between subtasks, edges are constructed between task nodes, and the edges between task nodes are endowed with corresponding attributes according to the execution order of the task nodes, and the association degree between the subtasks is determined according to the dependence degree and priority between the subtasks.

[0090] For example, according to the association degree between subtasks, edges are constructed between task nodes. For example, image acquisition task can be completed before image preprocessing task, and image preprocessing task can be completed before image recognition task. According to the dependence degree and priority between the subtasks, the association degree between the subtasks is determined, and edges are constructed between the task nodes, and the edges between the task nodes are endowed with corresponding attributes according to the execution order of the task nodes, such as the weight of the edge representing the closeness of the order of the tasks.

[0091] Step 2.4, according to the association degree between agents, edges are constructed between agent nodes, and the edges between agent nodes are endowed with corresponding attributes according to the coordination efficiency of the agents, and the association degree between the agents is determined according to the communication rate (such as bandwidth) between the agents and the related degree of processing tasks.

[0092] For example, according to the association degree between agents, edges are constructed between agent nodes. For example, agent A and agent B may have data interaction, and the association degree between the agents is determined according to the communication rate and the related degree of processing tasks between them, and edges are constructed between the agent nodes, and the edges between the agent nodes are endowed with corresponding attributes according to the coordination efficiency of the agents, such as the weight of the edge representing the coordination efficiency.

[0093] Step 2.5, according to the adaptation degree (such as core skill and task demand fit degree) between agents and subtasks, edges are constructed between task nodes and agent nodes, and the edges between task nodes and agent nodes are endowed with corresponding attributes.

[0094] Exemplarily, edges are constructed between the task nodes and the agent nodes according to the adaptation degrees between the agents and the subtasks, and the edges between the task nodes and the agent nodes are assigned with corresponding attributes. For example, if the adaptation degree between the agent A and the image recognition task is high, an edge is constructed between the image recognition task node and the agent A node, and the attribute of the edge can represent the adaptation score and the like.

[0095] Step 3, respectively according to the similarity degrees of the task adaptation abilities of the agents and the similarity degrees of the processing requirements of the to-be-scheduled tasks, the agent nodes and the task nodes in the heterogeneous interaction graph are clustered and grouped, and a heterogeneous interaction graph between the clustered and grouped agents and the to-be-scheduled tasks is obtained.

[0096] In an implementable manner, regarding the clustering and grouping of the agent nodes and the task nodes in the heterogeneous interaction graph according to the similarity degrees of the task adaptation abilities of the agents and the similarity degrees of the processing requirements of the to-be-scheduled tasks, and obtaining a heterogeneous interaction graph between the clustered and grouped agents and the to-be-scheduled tasks, the specific implementation includes agent node clustering, task node clustering and heterogeneous interaction graph reconstruction.

[0097] Regarding the agent node clustering, the specific implementation is as follows:

[0098] 1) The attributes of each agent node are extracted, abstracted into an agent feature vector, and the agent feature vector is standardized.

[0099] Exemplarily, the attributes of each agent node are extracted, such as the image recognition ability level of the agent A, the remaining amount of GPU resources, the network bandwidth and the like, and are abstracted into an agent feature vector.

[0100] 2) Based on the standardized agent feature vector, a similarity matrix between the agent nodes is constructed by using cosine similarity measurement.

[0101] In this embodiment, the similarity matrix between the agent nodes is constructed by using cosine similarity measurement, and the specific implementation is as follows:

[0102] Based on the agent matrix The similarity S between the agent nodes is calculated ij = v i · v j / ‖v i ‖·||v j ||;

[0103] Based on the similarity between the agent nodes, a similarity matrix between the agent nodes is constructed;

[0104] In the formula, represents the feature representation of the nth agent, and the dimension is h; S ijdenotes the similarity between agent node i and agent node j; v i denotes the feature vector of agent node i, with dimension h; v j denotes the feature vector of agent node j, with dimension h; ‖v i ‖denotes the norm of v i ; ||v j ||denotes the norm of v j .

[0105] 3) A label propagation clustering method is used to construct a weighted edge graph based on the original heterogeneous interaction graph using the similarity matrix between agent nodes.

[0106] 4) Each agent node in the weighted edge graph is assigned a unique and mutually exclusive initial label, and each agent node is independently grouped.

[0107] 5) The agent nodes in the weighted edge graph are traversed in a random order, and the label of each agent node is updated to the highest vote label using a weighted voting method based on neighbor labels and edge weights.

[0108] In this embodiment, the label of each agent node is updated to the highest vote label using a weighted voting method, specifically:

[0109]

[0110] where Score(c) denotes the number of votes for the current label of each agent node; W ij denotes the edge weight between the current agent node i and agent node j; j∈N(i) denotes only traversing the agent nodes directly connected to agent node i; L j =c denotes only counting the neighbor agent nodes j currently holding label c; the new label of agent node i is updated to

[0111] 6) Repeat step 5) until the maximum number of iterations is reached and the iteration is terminated when the label is completely stable, and finally aggregate the agent nodes with the same label into a group.

[0112] Regarding task node clustering, the following is specific:

[0113] 7) Extract the attributes of each task node and abstract them as task feature vectors, and perform standardization processing on the task feature vectors.

[0114] 8) Based on the standardized task feature vectors, a similarity matrix between task nodes is constructed using cosine similarity measurement.

[0115] In this embodiment, a similarity matrix between task nodes is constructed using cosine similarity measurement, specifically:

[0116] Based on the task matrix Calculate the similarity G between task nodes ij = u i · u j / ‖u i ‖·||u j ||;

[0117] Based on the similarity between task nodes, construct a similarity matrix between task nodes;

[0118] In the formula, denotes the feature representation of the nth task, with dimension h; G ij denotes the similarity between task node i and task node j; u i denotes the feature vector of task node i, with dimension h; u j denotes the feature vector of task node j, with dimension h;‖u i ‖denotes the norm of u i ;||u j ||denotes the norm of u j .

[0119] 9) Use the label propagation clustering method to construct a weighted edge graph based on the original heterogeneous interaction graph using the similarity matrix between task nodes.

[0120] 10) Assign a unique and mutually exclusive initial label to each task node in the weighted edge graph, and each task node is independent.

[0121] 11) Traverse the task nodes in the weighted edge graph in random order, use neighbor labels and edge weights, and use a weighted voting method to update the label of each task node to the highest vote label.

[0122] In this embodiment, the label of each task node is updated to the highest vote label using the weighted voting method, specifically:

[0123]

[0124] In the formula, Score(z) represents the number of votes for the current label of each task node; Q ij denotes the edge weight between the current task node i and task node j; j∈M(i) denotes only traversing the task nodes directly connected to task node i; L j =z denotes only counting neighbor task nodes j currently holding label z; the new label of task node i is updated to

[0125] 12) Repeat iteration step 11) until the maximum number of iterations is reached and the iteration is terminated when the labels are completely stable, finally the task nodes with the same label are aggregated into a group.

[0126] Regarding the heterogeneous interaction graph reconstruction, the following is specific:

[0127] 13) Abstract the agent nodes and task nodes belonging to the same group into super nodes respectively, and perform structural reconstruction on the basis of the original heterogeneous interaction graph to obtain the heterogeneous interaction graph between the clustered and grouped agents and the to-be-scheduled tasks.

[0128] Step 4, using the pre-trained multi-scale collaborative task scheduling model to perform coarse-grained processing on the heterogeneous interaction graph between the clustered and grouped agents and the to-be-scheduled tasks, to obtain a coarse-grained multi-scale collaborative task scheduling scheme, the coarse-grained multi-scale collaborative task scheduling scheme is a collaborative scheduling scheme between the clustered and grouped agents and the to-be-scheduled tasks.

[0129] That is, the clustered and grouped heterogeneous interaction graph is input into the multi-scale collaborative task scheduling model, and the multi-scale collaborative task scheduling model outputs the collaborative scheduling scheme between the clustered and grouped agents and the to-be-scheduled tasks, that is, the coarse-grained multi-scale collaborative task scheduling scheme, according to the input graph structure and node attribute information.

[0130] In this embodiment, the training process of the multi-scale collaborative scheduling model is as follows:

[0131] The agent-task collaboration dataset AirSim is used to train the graph neural network to obtain a multi-scale collaborative scheduling model capable of outputting agent-task matching relationships; the loss function used in the graph neural network training process comprehensively considers the task processing accuracy, task completion time and agent processing energy consumption, and by continuously adjusting the parameters of the graph neural network, the loss function is minimized, thereby obtaining a multi-scale collaborative scheduling model capable of outputting agent-task matching relationships.

[0132] The loss function is specifically:

[0133]

[0134] In the formula, k represents the kth task scheduling scheme; represents the time loss of the agent system executing the task; E ed represents the energy consumption of the agent e processing the task d; represents the accuracy loss of the agent system executing the task; represents the agent task allocation state, is 1 or 0, output in the multi-scale collaborative task arrangement scheme, 1 indicating that the task allocation is successful, and 0 indicating that the task is not allocated; a represents the weight of the accuracy loss, which can be set to 0.5 according to actual needs; β represents the weight of the time loss, which can be set to 0.3; and γ represents the weight of the energy consumption loss, which can be set to 0.2.

[0135] In the embodiment, the heterogeneous interaction graph between the clustered and grouped agents and the to-be-arranged tasks is coarsely processed by using the pre-trained multi-scale collaborative task arrangement model to obtain a coarse-grained multi-scale collaborative task arrangement scheme, and the specific process is as follows.

[0136] The heterogeneous interaction graph after clustering and grouping is input into the pre-trained multi-scale collaborative task arrangement model with the agent nodes and the task nodes as basic units, and the collaborative arrangement score between each pair of intelligent groups and task groups is output.

[0137] According to the collaborative arrangement score output by the model, a matching scheme of the coarse-grained intelligent groups and the task groups is generated to form a coarse-grained multi-scale collaborative task arrangement scheme as an input basis for subsequent fine-grained task allocation and scheduling.

[0138] Step 5, input the task node into the pre-constructed task scheduling priority model to obtain a task priority, and use the task priority to perform fine-grained processing on the coarse-grained multi-scale collaborative task arrangement scheme to obtain a fine-grained multi-scale collaborative task arrangement scheme, which is a collaborative arrangement scheme between the agents and the corresponding to-be-arranged tasks.

[0139] In the embodiment, the task scheduling priority model includes an optimization target and a constraint condition.

[0140] The optimization target is:

[0141]

[0142] The constraint condition is:

[0143] When two tasks are allocated to the same agent, it must be ensured that their execution times do not overlap.

[0144] In the formula, x d represents the start time of task d; y ed represents whether task d is placed on agent e; W d represents the priority of task d; C d represents the completion time of task d; T ed represents the execution time of task d on agent e; y ef represents whether task f is placed on agent e; T efdenotes the execution time of task f on agent e; f = argmin f＝d,k {x f} denotes the task with earlier start time between task d and task k; y ed = y ek = 1 indicates that task d and task k are assigned to the same agent e.

[0145] In the embodiment, regarding the fine-grained processing of the coarse-grained multi-scale collaborative task scheduling scheme by using the task priority, a fine-grained multi-scale collaborative task scheduling scheme is obtained, and the specific process is as follows.

[0146] Step 5.1, based on the matching relationship between the agent group and the task group determined in the coarse-grained multi-scale collaborative task scheduling scheme, the subtasks in the task group are mapped to the agents in the agent group, and a fine-grained allocation problem between the task subgraph and the agent subgraph is constructed.

[0147] Step 5.2, based on the task priority and in combination with the attributes of the agent nodes (such as the remaining amount of computing resources, communication performance, etc.), the subtasks in each task group are sorted to determine the allocation order. For example, tasks with high priority are preferentially allocated, and for tasks with the same priority, the agents are sorted from more to less according to the remaining amount of computing resources.

[0148] Step 5.3, based on the sorting result, the specific subtasks are allocated to the optimal agent in the agent group, while considering the currently available computing and communication resource capabilities and the current coordinate information in the physical or logical space. For example, an agent with sufficient computing resources and close to the task execution location is selected.

[0149] Step 5.4, after completing all fine-grained mapping, the final fine-grained multi-scale collaborative task scheduling scheme is output, which clearly indicates which agent executes each specific task, and can be directly input to the scheduling execution system to realize efficient collaborative task execution of the multi-agent system.

[0150] Through the above specific embodiments, the task processing capability of the heterogeneous agent system can be improved, the collaborative capability of the system can be enhanced, the system has good system scalability, and the problems existing in the prior art are effectively solved.

[0151] In another embodiment of the present application, a computer device is provided, which comprises a processor and a memory for storing a computer program comprising program instructions, and the processor is configured to execute the program instructions stored in the computer storage medium. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions in the computer storage medium to implement a corresponding method process or corresponding function; the processor in the embodiments of the present application can be used for the operation of the multi-scale task collaborative arrangement method for heterogeneous large language model multi-agent.

[0152] In another embodiment of the present application, the present application further provides a storage medium, specifically a computer readable storage medium (Memory), which is a memory device in a computer device, and is used for storing programs and data. It can be understood that the computer readable storage medium here can include the built-in storage medium in the computer device, and of course can also include the expansion storage medium supported by the computer device. The computer readable storage medium provides a storage space, which stores the operating system of the terminal. Moreover, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer readable storage medium here can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory. One or more instructions stored in the computer readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the multi-scale task collaborative arrangement method for heterogeneous large language model multi-agent in the above-mentioned embodiments.

[0153] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0154] The present application is described in reference to the flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0155] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0156] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0157] The application further provides a computer program product, which is used for executing any one of the above-mentioned multi-scale task collaborative arrangement methods for a heterogeneous large language model multi-agent. Since the computer program product provided by the application and the multi-scale task collaborative arrangement method for a heterogeneous large language model multi-agent belong to the same inventive concept, the computer program product provided by the application has all the advantages of the multi-scale task collaborative arrangement method for a heterogeneous large language model multi-agent, and thus the beneficial effects of the computer program product provided by the application will not be described one by one.

[0158] In the present application, the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" mean that the specific feature, structure, material or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the different embodiments or examples described in the present specification and the features of the different embodiments or examples can be combined and combined by those skilled in the art without contradiction.

[0159] Finally, it should be noted that: the above-described embodiments are only specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit them, the protection scope of the present application is not limited thereto, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments within the technical range disclosed by the present application, or make equivalent replacement to some technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and all should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A multi-scale task coordination arrangement method for heterogeneous large language model multi-agent, characterized in that, The method comprises the following steps: obtaining processing requirements of a task to be arranged and task adaptation capabilities of each agent; constructing a heterogeneous interaction graph between the agents and the task to be arranged according to the processing requirements of the task to be arranged and the task adaptation capabilities of each agent, wherein the heterogeneous interaction graph comprises agent nodes and task nodes; clustering and grouping the agent nodes and the task nodes in the heterogeneous interaction graph according to the similarity of the task adaptation capabilities of the agents and the similarity of the processing requirements of the task to be arranged, to obtain a clustered and grouped heterogeneous interaction graph between the agents and the task to be arranged; performing coarse-grained processing on the clustered and grouped heterogeneous interaction graph between the agents and the task to be arranged by using a pre-trained multi-scale collaborative task arrangement model, to obtain a coarse-grained multi-scale collaborative task arrangement scheme, wherein the coarse-grained multi-scale collaborative task arrangement scheme is a collaborative arrangement scheme between the agents and the task to be arranged; inputting the task nodes into a pre-constructed task scheduling priority model to obtain a task priority, and performing fine-grained processing on the coarse-grained multi-scale collaborative task arrangement scheme by using the task priority, to obtain a fine-grained multi-scale collaborative task arrangement scheme, wherein the fine-grained multi-scale collaborative task arrangement scheme is a collaborative arrangement scheme between the agents and the corresponding task to be arranged.

2. The multi-scale task collaborative arrangement method for heterogeneous large language model multi-agent according to claim 1, characterized in that, The processing requirements of the task to be arranged comprise computing resource requirements, storage requirements and urgency requirements. The task adaptation capabilities of the agents comprise core capabilities of the agents for completing specific tasks, currently available computing and communication resource capabilities, and current coordinate information in physical or logical space.

3. The multi-scale task coordination method for heterogeneous large language model multi-agent according to claim 1, characterized in that, The construction of the heterogeneous interaction graph between the agents and the task to be arranged according to the processing requirements of the task to be arranged and the task adaptation capabilities of each agent comprises the following steps: decomposing the task to be arranged into a plurality of subtasks according to the processing requirements of the task to be arranged, abstracting each subtask into a task node, and assigning corresponding attributes to each task node according to the processing requirements of each subtask; abstracting each agent into an agent node, and assigning corresponding attributes to the corresponding agent node according to the task adaptation capabilities of each agent; constructing edges between the task nodes according to the association degree between the subtasks, and assigning corresponding attributes to the edges between the task nodes according to the execution order of the task nodes, wherein the association degree between the subtasks is determined according to the dependency degree and priority between the subtasks; constructing edges between the agent nodes according to the association degree between the agents, and assigning corresponding attributes to the edges between the agent nodes according to the collaborative efficiency of the agents, wherein the association degree between the agents is determined according to the communication rate between the agents and the correlation degree of processing tasks; constructing edges between the task nodes and the agent nodes according to the adaptation degree between the agents and the subtasks, and assigning corresponding attributes to the edges between the task nodes and the agent nodes.

4. The multi-scale task coordination arrangement method for heterogeneous large language model multi-agent according to claim 1, characterized in that, The agent nodes and task nodes in the heterogeneous interaction graph are clustered and grouped according to the similarity of the task adaptation capabilities of the agents and the similarity of the processing requirements of the tasks to be arranged, and a heterogeneous interaction graph between the clustered and grouped agents and the tasks to be arranged is obtained, specifically comprising: 1) Extract the attributes of each agent node, abstract them as agent feature vectors, and standardize the agent feature vectors; 2) Based on the standardized agent feature vectors, a similarity matrix between agent nodes is constructed using cosine similarity measurement; 3) Using a label propagation clustering method, a weighted edge graph is constructed based on the original heterogeneous interaction graph using the similarity matrix between agent nodes; 4) Each agent node in the weighted edge graph is assigned a unique and mutually exclusive initial label, and each agent node is independently grouped; 5) Traverse the agent nodes in the weighted edge graph in random order, and use neighbor labels and edge weights to update the label of each agent node to the highest vote label using a weighted voting method; 6) Repeat step 5) until the maximum number of iterations is reached and the label is completely stable, and finally aggregate agent nodes with the same label into a group; 7) Extract the attributes of each task node, abstract them as task feature vectors, and standardize the task feature vectors; 8) Based on the standardized task feature vectors, a similarity matrix between task nodes is constructed using cosine similarity measurement; 9) Using a label propagation clustering method, a weighted edge graph is constructed based on the original heterogeneous interaction graph using the similarity matrix between task nodes; 10) Each task node in the weighted edge graph is assigned a unique and mutually exclusive initial label, and each task node is independently grouped; 11) Traverse the task nodes in the weighted edge graph in random order, and use neighbor labels and edge weights to update the label of each task node to the highest vote label using a weighted voting method; 12) Repeat step 11) until the maximum number of iterations is reached and the label is completely stable, and finally aggregate task nodes with the same label into a group; 13) Abstract the agent nodes and task nodes belonging to the same group as super nodes, and reconstruct the structure based on the original heterogeneous interaction graph to obtain the clustered and grouped heterogeneous interaction graph between the agents and the tasks to be arranged.

5. The multi-scale task coordination method for heterogeneous large language model multi-agent according to claim 4, characterized in that, The similarity matrix between agent nodes is constructed using cosine similarity measurement, specifically: Agent matrix based Computing similarity S between agent nodes ij = v i · v j / ‖ v i ‖·|| v j || The similarity matrix between agent nodes is constructed based on the similarity between agent nodes; wherein, represents the feature representation of the nth agent, with dimension h; S ij represents the similarity between agent node i and agent node j; v i represents the feature vector of agent node i, with dimension h; v j represents the feature vector of agent node j, with dimension h; ‖v i ‖ represents the norm of v i ; ||v j || represents the norm of v j ; The similarity matrix between task nodes is constructed using cosine similarity measurement, specifically: Based on a task matrix Computing similarity G between task nodes ij = u i · u j / ‖u i ‖·||u j || The similarity matrix between task nodes is constructed based on the similarity between task nodes; wherein, represents the feature representation of the nth task, with dimension h; G ij represents the similarity between task node i and task node j; u i represents the feature vector of task node i, with dimension h; u j represents the feature vector of task node j, with dimension h; ‖u i represents the norm of u i ; ||u j || represents the norm of u j .

6. The multi-scale task coordination method for heterogeneous large language model multi-agent according to claim 4, characterized in that, The label of each agent node is updated to the highest vote label using a weighted voting method, specifically: wherein Score(c) represents the number of votes of the current label of each agent node; E ij represents the edge weight between the current agent node i and agent node j; j∈N(i) represents only traversing the agent nodes directly connected to the agent node i; L j =c represents only counting the neighbor agent node j currently holding the label c; the new label of the agent node i is updated as The label of each task node is updated to the highest vote label using a weighted voting method, specifically: wherein Score(z) represents the vote number of the current label of each task node; Q ij represents the edge weight between the current task node i and task node j; j ∈ M(i) represents only traversing the task nodes directly connected to task node i; L j =z represents only counting the neighbor task nodes j currently holding label z; the new label of task node i is updated as 7. The multi-scale task coordination method for heterogeneous large language model multi-agent according to claim 1, characterized in that, The training process of the multi-scale collaborative scheduling model includes: The graph neural network is trained using an agent-task collaboration dataset AirSim to obtain a multi-scale collaborative scheduling model capable of outputting an agent-task matching relationship; a loss function used in the training process of the graph neural network comprehensively considers task processing accuracy, task completion time, and agent processing energy consumption, and the loss function is specifically: In the formula, k represents the kth task scheduling scheme; represents the time loss of the agent system executing the task; E ed represents the energy consumption of the agent e processing the task d; represents the accuracy loss of the agent system executing the task; represents the agent task allocation state, The value of is 1 or 0, which is output in the multi-scale collaborative task scheduling scheme, 1 represents that the task allocation is successful, and 0 represents that the task is not allocated; α represents the weight of the accuracy loss; β represents the weight of the time loss; γ represents the weight of the energy consumption loss.

8. The multi-scale task coordination method for heterogeneous large language model multi-agent according to claim 1, characterized in that, The task scheduling priority model includes an optimization objective and constraint conditions: The optimization objective is: Constraint: |x d - x k | ≥ y ef · T ef , y ed = y ek = 1, f = argmin f＝d,k {x f} where x d denotes the start time of task d; y ed denotes whether task d is placed on agent e; W d denotes the priority of task d; C d denotes the completion time of task d; t ed denotes the execution time of task d on agent e; y ef denotes whether task f is placed on agent e; T ef denotes the execution time of task f on agent e; f = argmin f＝d,k {x f} denotes the task with earlier start time among task d and task k; y ed = y ek = 1 denotes that task d and task k are assigned to the same agent e.

9. The multi-scale task coordination method for heterogeneous large language model multi-agent according to claim 1, characterized in that, The task scheduling priority model includes an optimization objective and constraint conditions: The task scheduling priority model includes an optimization objective and constraint conditions: Based on the matching relationship between the agent groups and the task groups determined in the coarse-grained multi-scale collaborative task scheduling scheme, the subtasks in the task groups are mapped to the agents in the agent groups, and a fine-grained allocation problem between the task subgraphs and the agent subgraphs is constructed; Based on the task priority and in combination with the attributes of the agent nodes, the subtasks in each task group are sorted; Based on the sorting result, the specific subtasks are allocated to the optimal agents in the agent groups, while considering the currently available computing and communication resource capabilities and the current coordinate information in the physical or logical space; 10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, After completing all fine-grained mapping, the final fine-grained multi-scale collaborative task scheduling scheme is output. The processor executes the computer program to implement the multi-scale task collaborative scheduling method for heterogeneous large language model multi-agents according to any one of claims 1 to 9.

Images