Multi-agent cooperative optimization method and system based on multi-terminal training access

Abstract

The application relates to the technical field of agent cooperation optimization, and discloses a multi-agent cooperation optimization method and system based on multi-terminal training access, which comprises the following steps: constructing an operation intention vector according to multi-modal training operation data from different terminals; mapping the operation intention vector to a pre-constructed task graph to predict a first computing resource; setting a coupling relationship between each agent and a task in different task roles, calculating the value of an agent combination executing a task sequence, and obtaining a first resource allocation strategy; simulating a task execution process of the agent combination executing the task sequence, optimizing the first resource allocation strategy, and obtaining a second resource allocation strategy; and allocating corresponding computing resources to each agent according to the second resource allocation strategy, and scheduling each agent to execute each task in the task sequence. The application can improve resource utilization, improve agent cooperation performance, and thus improve the quality and efficiency of multi-terminal training.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agent cooperation optimization, and more particularly to a multi-agent cooperation optimization method and system based on multi-terminal training access. BACKGROUND

[0002] At present, with the popularization of multi-terminal access training mode, the system needs to process multi-modal operation data from different sources and coordinate multiple agents to complete complex training tasks. The existing multi-agent cooperation scheme usually adopts a preset task flow and a static resource allocation strategy, which is difficult to dynamically and accurately understand the operation intention of the user, and cannot adaptively optimize resources and fault-tolerant scheduling according to real-time task requirements and agent characteristics, resulting in low training efficiency, insufficient resource utilization and inability to adapt to complex training scenarios.

[0003] The existing technology has the following problems: the operation data of different terminals are analyzed separately, there is a lack of joint analysis of different terminal data, and it is difficult to accurately understand the real operation intention of the user; based on a fixed task decomposition process, the association between tasks is not considered, resulting in inaccurate resource prediction results and inability to dynamically analyze tasks; based on a single agent resource allocation strategy to control the agent cooperation process, it cannot adapt to the influence of dynamic agent combination on system resource utilization, resulting in low system performance; to solve at least one of the above problems, the present application proposes a multi-agent cooperation optimization method and system based on multi-terminal training access. SUMMARY

[0004] In view of the deficiencies of the prior art, the purpose of the present application is to provide a multi-agent cooperation optimization method and system based on multi-terminal training access, which can effectively solve the problems in the background art. The specific technical solutions of the present application are as follows:

[0005] The multi-agent cooperation optimization method based on multi-terminal training access comprises:

[0006] According to the multi-modal training operation data from different terminals, the operation action, operation entity and corresponding constraint condition are extracted through context association analysis to construct an operation intention vector;

[0007] The operation intention vector is mapped into a pre-constructed task graph to identify a task role decomposition to obtain a task sequence, analyze the historical performance data of the corresponding tasks in the task sequence, and predict a first computing resource;

[0008] The coupling relationship between each agent and the tasks in different task roles is set, the value of the corresponding agent combination executing the task sequence is calculated, the first computing resource is allocated to the corresponding agent according to the value, and a first resource allocation strategy is obtained;

[0009] According to the first resource allocation strategy, a simulated agent combines to perform a task execution process of the task sequence, identifies a fault link in the task execution process by adding simulated interference, optimizes the first resource allocation strategy, and obtains a second resource allocation strategy;

[0010] According to the second resource allocation strategy, each agent is allocated corresponding computing resources, and each agent is scheduled to perform each task in the task sequence, so as to optimize the multi-agent collaboration process.

[0011] Specifically, the operation intention vector is constructed according to the multi-modal training operation data from different terminals, through context association analysis to extract operation actions, operation entities and corresponding constraint conditions, including:

[0012] According to the multi-modal training operation data from different terminals, the pause of the operation action and the switching rule of the teaching link are analyzed, the operation process boundary point is identified, the multi-modal training operation data is divided into a plurality of operation segments, and an operation segment set is obtained;

[0013] Based on the operation segment set, feature extraction is performed through a feature extraction model to obtain a feature sequence;

[0014] According to the feature sequence, the interaction state and the operation tool usage of the operation process are analyzed, the mapping relationship between the operation entity and the training knowledge is constructed, and the operation entity association matrix is calculated;

[0015] In the operation entity association matrix, the abnormal behavior mode and the operation habit of different users are analyzed, the time constraint and the abnormal condition are extracted, and a constraint condition set is constructed;

[0016] Combined with the operation segment set, the operation entity association matrix and the constraint condition set, the teaching structure, the operation association and the constraint priority are analyzed, and the operation intention vector is constructed.

[0017] Specifically, the operation intention vector is mapped into a pre-constructed task graph, a task sequence is identified by decomposing a task role, historical performance data of corresponding tasks in the task sequence are analyzed, and a first computing resource is predicted, including:

[0018] A dynamic subgraph matching mechanism is configured, the operation intention vector is mapped into a pre-constructed task graph, a task sequence is identified by decomposing a task role, and the operation intention vector is mapped into a pre-constructed task graph.

[0019] According to the task sequence, the historical performance data corresponding to each task is analyzed, the task features are extracted, and the first performance data of each task is calculated;

[0020] Analyze the dependency relationship between tasks, identify the critical path in the task sequence according to the dependency relationship, combine the task resource demand condition of the critical path and the resource overlapping competition condition of the non-critical path, and predict the first computing resource that meets the computing demand of the task sequence.

[0021] Specifically, the dynamic subgraph matching mechanism is configured to map the operation intention vector into the pre-constructed task graph, identify the task sequence by decomposing the task role, including:

[0022] According to the operation intention vector, identify the conflict point of different terminal operation intentions at the same time sequence position, and perform intention deduplication or intention fusion at the conflict point to obtain a first intention vector;

[0023] Based on the first intention vector, analyze the distance between the operation entity and the task node in the pre-constructed task graph, calculate the intention response degree of each task node, and select the task nodes with an intention response degree greater than a preset response threshold as a first task node set;

[0024] In the first task node set, analyze the structural similarity between the operation intention and the task node execution template, and identify the task deviation;

[0025] Combine the task deviation and the weight of the intelligent agent corresponding to each task node to calculate the weight of the intelligent agent corresponding to each task node, and assign the task to the corresponding intelligent agent to obtain a task sequence.

[0026] Specifically, the coupling relationship between each intelligent agent and the tasks in different task roles is set, the value of the corresponding intelligent agent combination executing the task sequence is calculated, the first computing resource is allocated to the corresponding intelligent agent according to the value, and a first resource allocation strategy is obtained, including:

[0027] Analyze the adaptation degree of different task roles and corresponding task node intelligent agents in the task sequence, construct a coupling relationship, analyze the resource change of each task node in the terminal switching process according to the coupling relationship, and calculate the task execution priority;

[0028] According to the task execution priority from high to low, simulate the value of different intelligent agent combinations executing the task sequence, and allocate the first computing resource to the corresponding intelligent agent according to the value to obtain a first resource allocation strategy.

[0029] Specifically, according to the task execution priority from high to low, simulate the value of different intelligent agent combinations executing the task sequence, and allocate the first computing resource to the corresponding intelligent agent according to the value to obtain a first resource allocation strategy, including:

[0030] According to the task execution priority from high to low, simulate the concurrent execution process of different intelligent agent combinations at each task node, analyze the resource consumption and teaching effect in the execution process, and calculate the value of the execution process.

[0031] Based on the value, the marginal benefit of resource demand in the first computing resource to different agents is analyzed, and a resource mapping relationship of the agents is constructed in combination with the resource demand of each agent under different terminals.

[0032] According to the resource mapping relationship of the agents, the first computing resource is allocated to the corresponding agents to obtain a first resource allocation strategy.

[0033] Specifically, according to the first resource allocation strategy, a task execution process of the agent combination executing the task sequence is simulated, a failure link in the task execution process is identified by adding a simulated interference, the first resource allocation strategy is optimized, and a second resource allocation strategy is obtained, including:

[0034] According to the first resource allocation strategy, a task execution process of the agent combination executing the task sequence is simulated, a failure link in the task execution process is identified by adding a simulated interference, the first resource allocation strategy is optimized, and a second resource allocation strategy is obtained, including:

[0035] According to the failure link, the correlation between different failure types and resource allocation is analyzed, the first resource allocation strategy is optimized, and a second resource allocation strategy is obtained.

[0036] Specifically, according to the first resource allocation strategy, a task execution process of the agent combination executing the task sequence is simulated, a failure link in the task execution process is identified by adding a simulated interference, the first resource allocation strategy is optimized, and a second resource allocation strategy is obtained, including:

[0037] According to the first resource allocation strategy, a task execution process of the agent combination executing the task sequence is simulated, a failure link in the task execution process is identified by adding a simulated interference, the first resource allocation strategy is optimized, and a second resource allocation strategy is obtained, including:

[0038] Based on the first execution result, an interference mode corresponding to a multi-terminal training process is extracted, the adaptation degree of each interference mode to the training process is calculated, and an interference matrix is constructed.

[0039] According to the interference matrix, corresponding interference is added in the task execution simulation process, the failure condition is analyzed, and the failure link is identified.

[0040] Specifically, according to the second resource allocation strategy, corresponding computing resources are allocated to each agent, and each agent is scheduled to execute each task in the task sequence, so as to optimize the multi-agent cooperation process, including:

[0041] According to the second resource allocation strategy, corresponding computing resources are allocated to each agent.

[0042] In combination with the computing resource allocation and the task execution priority, each agent is scheduled to execute each task in the task sequence, so as to optimize the multi-agent cooperation process.

[0043] The multi-agent cooperation optimization system based on multi-terminal training access is used to realize the multi-agent cooperation optimization method based on multi-terminal training access, and comprises:

[0044] An operation intention analysis module extracts operation actions, operation entities and corresponding constraint conditions through context association analysis according to multi-modal training operation data from different terminals, and constructs an operation intention vector.

[0045] A computing resource prediction module maps the operation intention vector into a pre-constructed task graph, identifies task role decomposition to obtain a task sequence, analyzes historical performance data of corresponding tasks in the task sequence, and predicts a first computing resource.

[0046] A resource allocation module sets a coupling relationship between each agent and a task in different task roles, calculates the value of the corresponding agent combination executing the task sequence, allocates the first computing resource to the corresponding agent according to the value, and obtains a first resource allocation strategy.

[0047] A resource optimization module simulates the task execution process of the agent combination executing the task sequence according to the first resource allocation strategy, identifies fault links in the task execution process by adding simulated interference, optimizes the first resource allocation strategy, and obtains a second resource allocation strategy.

[0048] A multi-agent cooperation module allocates corresponding computing resources to each agent according to the second resource allocation strategy, schedules each agent to execute each task in the task sequence, and optimizes the multi-agent cooperation process.

[0049] The application has the following beneficial effects: The operation actions, operation entities and constraint conditions in the multi-modal training operation data are associated and analyzed, the operation intention vector is constructed, the operation intention vector is mapped into the task graph through the dynamic subgraph matching mechanism, the intention conflict problem in the multi-terminal concurrent operation can be avoided, the corresponding task sequence is generated, the coupling between the agent and the task role is performed, the value distribution calculation resource of the task execution process of different agent combinations is simulated, the system performance is improved through dynamic resource allocation, the task execution process of the agent is simulated and various simulated interference is actively injected, the fault links are identified to optimize the resource allocation strategy; The accurate user operation intention is used for computing resource prediction and allocation, which can improve the resource utilization rate, the simulated interference optimizes the resource strategy, which can effectively cope with various abnormal situations in the complex operating environment, improves the agent cooperation performance, and thus improves the quality and efficiency of the multi-terminal training. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The workflow diagram of the multi-agent cooperation optimization method based on multi-terminal training access in the embodiments of the application;

[0051] Figure 2 A schematic diagram of a task dependency network graph in an embodiment of the present application;

[0052] Figure 3 A workflow diagram of a first resource allocation strategy calculation process in an embodiment of the present application;

[0053] Figure 4 A structural schematic diagram of a multi-agent collaborative optimization system based on multi-terminal training access in an embodiment of the present application. DETAILED DESCRIPTION

[0054] The present application will be further described below in conjunction with the drawings and embodiments.

[0055] In the embodiments of the present application, the words such as "exemplary" or "for example" are used to mean serving as an example, instance, or illustration. Any embodiment or design presented as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Rather, the use of the words such as "exemplary" or "for example" is intended to present concepts in a concrete manner.

[0056] Hereinafter, the terms "first", "second", and the like are used in a general sense and only for the purpose of description, and should not be construed or implied to indicate or imply relative importance or implicitly indicate the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more features. In the description of the embodiments of the present application, the meaning of "a plurality of" is two or more, unless otherwise specified.

[0057] Reference Figure 1 As shown, the specific embodiments of the multi-agent collaborative optimization method based on multi-terminal training access of the present application include:

[0058] S101, according to the multi-modal training operation data from different terminals, operation actions, operation entities and corresponding constraint conditions are extracted through context association analysis, and an operation intention vector is constructed;

[0059] S102, mapping the operation intention vector into a pre-constructed task graph, identifying task role decomposition to obtain a task sequence, analyzing the historical performance data of the corresponding tasks in the task sequence, and predicting a first computing resource;

[0060] S103, setting the coupling relationship between each agent and the tasks in different task roles, calculating the value of the corresponding agent group executing the task sequence, allocating the first computing resource to the corresponding agent according to the value, and obtaining a first resource allocation strategy;

[0061] S104, simulate the task execution process of the task sequence by the agent group according to the first resource allocation strategy, identify the fault link in the task execution process by adding simulated interference, optimize the first resource allocation strategy to obtain a second resource allocation strategy;

[0062] S105, allocate corresponding computing resources to each agent according to the second resource allocation strategy, and schedule each agent to execute each task in the task sequence, to optimize the multi-agent collaboration process.

[0063] With the deep integration of mixed reality, remote collaboration and intelligent teaching system, modern vocational training scene has gradually evolved into a complex distributed system architecture with multi-terminal heterogeneous access, multi-modal interactive collaboration and multi-agent service support. Students participate in training through various terminals such as touch tablets, voice assistants, somatosensory devices and AR glasses, generating high-concurrency and high-heterogeneous operation instruction streams; the system needs to schedule teaching guide agents, operation evaluation agents, scene rendering agents and other specialized agents to collaboratively complete knowledge transmission, skill practice and real-time feedback loop.

[0064] In this embodiment, according to multi-modal training operation data from different terminals, including but not limited to operation logs, video streams and sensor data, operation actions, operation entities and corresponding constraint conditions are extracted through context association analysis to construct operation intention vectors; unstructured raw data is converted into intention vectors that can be accurately understood and processed by constructing operation intention vectors, providing accurate data basis for task planning, and by introducing constraint conditions, the safety boundary and rule limit of operation can be analyzed, the accuracy and compliance of task decomposition and resource scheduling are improved, and the problem of collaboration process failure caused by intention misjudgment is reduced.

[0065] Specifically, the conflict points of operation intentions of different terminals on the timeline are identified through a dynamic subgraph matching mechanism, and the first intention vector is obtained by intention deduplication or intention fusion of the conflict points, the operation intention vector is mapped to the pre-constructed task graph according to the first intention vector, the task role decomposition is identified to obtain the task sequence, and the historical performance data of the corresponding task in the task sequence are analyzed, including but not limited to average execution time, CPU / memory consumption peak, to predict the first computing resource; the intention conflict problem in multi-terminal collaboration is effectively solved through the dynamic subgraph matching and intention fusion mechanism, the logical consistency and feasibility of the task sequence are guaranteed, and the accuracy of the resource prediction result is improved by combining dynamic task topology and historical performance to calculate and predict, which provides an accurate data basis for resource allocation, avoids the blindness of resource allocation, and improves the efficiency of resource allocation process and agent collaboration process.

[0066] Specifically, the matching degree between the inherent capability attribute of the agent and the required capability of the task role is analyzed, the coupling relationship between each agent and the tasks in different task roles is set, the process of combining and concurrently executing the task sequence by different agents is simulated, the value of executing the task sequence by the corresponding agent combination is calculated, the first computing resource is allocated to the corresponding agent according to the value, and the first resource allocation strategy is obtained; by analyzing the coupling relationship between the agent and the task and allocating resources based on the value, the limited computing resources are accurately allocated to the key link that can most improve the overall collaboration efficiency and the final teaching effect, the utilization efficiency of resources is improved, the high-priority tasks and core agents can obtain sufficient resource guarantee, and the performance and collaboration efficiency of the multi-agent collaboration process are improved.

[0067] According to the first resource allocation strategy, the task execution process of the agent combination executing the task sequence is simulated, the first execution result is obtained, the fault link in the task execution process is identified by adding simulated interference in the task execution process, the fault link is attributed, the association between the fault type and the resource allocation situation is established, the first resource allocation strategy is optimized, and the second resource allocation strategy is obtained; by simulating the fault interference, the fault point of the resource allocation strategy is identified before the fault occurs, and the strategy optimization is performed on the fault point, which can reduce the probability of performance degradation or even task failure of the system in actual operation due to unpredictable interference, and improve the stability and reliability of the system in a complex and dynamic real environment.

[0068] Specifically, each agent is allocated corresponding computing resources according to the second resource allocation strategy, and each agent executes each task in the task sequence to optimize the multi-agent collaboration process; based on the optimized strategy, the resources are allocated and the tasks are scheduled, which ensures that the entire multi-agent collaboration system can stably run under the premise of sufficient resources and reasonable configuration, optimizes the multi-agent collaboration process, and improves the task execution efficiency, resource utilization rate and system running performance.

[0069] The application analyzes the operation actions, operation entities and constraint conditions in the multi-modal training operation data, constructs an operation intention vector, maps the operation intention vector to a task graph through a dynamic subgraph matching mechanism, can avoid intention conflict problems when multiple terminals operate concurrently, generates a corresponding task sequence, couples the agent and the task role, simulates the value distribution calculation resource of the task execution process of different agent combinations, improves the system performance through dynamic resource allocation, simulates the task execution process of the agent and actively injects various simulation interferences, identifies the fault link to optimize the resource allocation strategy; based on the accurate user operation intention, the calculation resource prediction and allocation can improve the resource utilization rate, the simulation interference optimizes the resource strategy, can effectively cope with various abnormal situations in the complex operating environment, improves the agent cooperation performance, thereby improving the quality and efficiency of multi-terminal training.

[0070] Further, according to the multi-modal training operation data from different terminals, the operation actions, operation entities and corresponding constraint conditions are extracted through context association analysis, and an operation intention vector is constructed, including:

[0071] S201, according to the multi-modal training operation data from different terminals, the operation action pause and teaching link switching rule are analyzed, the operation process boundary point is identified, the multi-modal training operation data is divided into multiple operation segments, and an operation segment set is obtained;

[0072] S202, based on the operation segment set, feature extraction is performed through a feature extraction model to obtain a feature sequence;

[0073] S203, according to the feature sequence, the interaction state and operation tool usage of the operation process are analyzed, the mapping relationship between the operation entity and the training knowledge is constructed, and the operation entity association matrix is calculated;

[0074] S204, in the operation entity association matrix, the abnormal behavior mode and the operation habit of different users are analyzed, the time constraint and the abnormal condition are extracted, and a constraint condition set is constructed;

[0075] S205, combining the operation segment set, the operation entity association matrix and the constraint condition set, the teaching structure, the operation association and the constraint priority are analyzed, and an operation intention vector is constructed.

[0076] In this embodiment, according to the multi-modal training operation data from different terminals, including but not limited to mouse click stream, touch gesture sequence, VR controller motion trajectory, voice instruction stream data, through the time sequence analysis algorithm, the pause interval in the operation data stream from each terminal is monitored, which occurs between two operation actions, and the preset teaching link switching signal is detected, including but not limited to the page turning of courseware, the completion mark of a stage in the training simulator, and the specific control instruction issued by the instructor terminal. Combined with the statistical characteristics of the pause interval and the teaching link switching event, the time when the pause interval and the teaching link switching occur at the same time is recognized as the operation process boundary point. The multi-modal data stream is cut at the operation process boundary point position, and the data is divided into multiple operation segments to obtain an operation segment set. By recognizing the operation boundary and cutting the data, the operation segment is more consistent with the teaching process logic, each operation segment corresponds to an operation process, improves the accuracy and efficiency of the subsequent analysis process, avoids feature confusion and intention misjudgment caused by mismatch between data and semantic boundary, and improves the accuracy of the analysis process.

[0077] Specifically, based on the operation segment set, feature extraction is performed through a feature extraction model. For video data, a pre-trained convolutional neural network is used to extract spatial and temporal features from each frame. For time series data stream, a pre-trained recurrent neural network is used to analyze the sequential dependence of operation actions and extract features with temporal context information. The features of each operation segment are extracted respectively, and the features are arranged in order according to the time sequence of the operation segments to obtain a feature sequence. Through feature extraction, the data amount can be compressed and the information density can be improved to obtain pattern information related to the identification of operation intention, and the accuracy and efficiency of cross-modal correlation analysis are improved.

[0078] According to the feature sequence, the interaction state and operation tool usage of the operation process are analyzed, and the corresponding operation entity of the current operation is identified, including but not limited to specific tools, equipment components or virtual interface elements in maintenance training; by querying the pre-constructed training knowledge graph according to a large amount of historical operation data, the identified operation entity is associated with the graph node, and the semantic information such as attributes and functions of the corresponding entity is extracted, the interaction state between entities within the time range covered by the feature sequence is analyzed, the co-occurrence frequency and operation sequence between entities are counted, and the correlation strength between entities in the feature sequence is analyzed based on the attention mechanism. An operation entity correlation matrix is calculated, the rows and columns of the matrix are operation entities, and the element values in the matrix represent the correlation strength of any two entities in the operation process.

[0079] It should be noted that the quantitative analysis of the association between the operation entities can reflect the logical and semantic structure between the entities in the operation process, and the calculated operation entity association matrix can reflect the core element network involved in the current operation, providing accurate data basis for analyzing the operation intention of multiple entities. Compared with the recognition of a single action, the association between actions can be analyzed, and the accuracy and efficiency of the user intention analysis result can be improved.

[0080] Specifically, in the operation entity association matrix, the abnormal behavior mode and the operation habits of different users are analyzed. For the abnormal behavior mode, the operation entity association matrix is compared with the ideal association matrix corresponding to the standard operation process stored in the knowledge base, and the association combination deviating from the normal mode is screened out. For the operation habits of different users, the entity association mode repeatedly appearing in the historical operation data of the same user is clustered and analyzed, and the corresponding operation habit is identified. Based on the identification result of the abnormal behavior mode, the warning or correction condition triggered when the abnormal entity association occurs is screened out to obtain the abnormal condition. The operation sequence and the operation entity association matrix are analyzed in time sequence combined with the operation habits of the user, the operation order between entities is extracted, the time constraint is constructed, and the constraint condition set is obtained combined with the time constraint and the abnormal condition. By extracting the time constraint and the abnormal condition, the rule boundary that must be followed in the operation process can be identified, which provides data basis for compliance judgment in the process of intention understanding and task planning. By identifying the user habits, the style of different operators can be adapted, and the adaptability to different users and environments can be improved.

[0081] Specifically, by using the graph neural network encoder pre-trained with a large amount of historical data, combining the operation fragment set, the operation entity association matrix and the constraint condition set, the graph neural network encoder analyzes the teaching structure, understands the stage and purpose of the current operation in the entire teaching outline or task flow, analyzes the operation association by judging the operation action and entity association, analyzes the constraint priority by judging the order of safety constraints, and the graph neural network encoder fuses the analysis results and maps them to a low-dimensional vector space to obtain an operation intention vector. By analyzing the operation intention vector, the multi-modal, multi-step and multi-entity user operation is converted into a structured data object, and the accuracy and efficiency of the task graph mapping analysis process are improved.

[0082] Further, the operation intention vector is mapped into a pre-constructed task graph, a task sequence is identified by decomposing the task role, and historical performance data of corresponding tasks in the task sequence is analyzed to predict a first computing resource, including:

[0083] S301, a dynamic subgraph matching mechanism is configured, the operation intention vector is mapped into a pre-constructed task graph, and a task sequence is identified by decomposing the task role;

[0084] S302, according to the task sequence, analyze the historical performance data corresponding to each task, extract task features, and calculate the first performance data of each task;

[0085] S303, analyze the dependency relationship between tasks, identify the critical path in the task sequence according to the dependency relationship, and combine the task resource demand situation of the critical path and the resource overlapping competition situation of the non-critical path to predict the first computing resource that meets the computing demand of the task sequence.

[0086] In this embodiment, a dynamic subgraph matching mechanism is configured to map the operation intention vector to the pre-constructed task graph, identify the task role decomposition, and obtain the task sequence. The dynamic subgraph matching mechanism can effectively solve the intention conflict problem in multi-user concurrent operation, guarantee the logical consistency and feasibility of the task sequence, and generate a task sequence that meets the actual operation intention of the user, thereby improving the accuracy and flexibility of task decomposition.

[0087] Specifically, according to the task sequence, the performance records of the task in past multiple execution instances are retrieved from the historical performance database, including but not limited to task execution time, CPU occupancy, memory consumption, network I / O amount, etc. Task features are extracted from the historical performance data, including but not limited to the average value, peak value, and fluctuation variance of the performance indicators. By using a random forest model pre-trained with a large amount of historical performance data, the historical features are analyzed in combination with the initial state of the current system to calculate the first performance data required for executing the task in the current expected environment. By analyzing the historical performance data and extracting key features, accurate analysis and prediction can be performed based on actual running data. The calculated first performance data conforms to the performance of the task in the real environment and can reflect the resource demand peak and typical demand of the task under pressure, thereby providing accurate data support for resource planning and improving the accuracy of resource estimation.

[0088] As Figure 2As shown, the dependency relationship between the analysis tasks is analyzed, a task dependency network diagram is constructed, the total time consumption of all possible paths is calculated according to the first performance data of each task, and the longest path is taken as a critical path; the resource requirements of all tasks on the critical path are summarized, the tasks on the non-critical path are analyzed, the tasks on the non-critical path and the critical path tasks will exist overlap in time, which will cause resource overlapping competition for the shared resource pool, the resource waiting loss caused by the resource overlapping competition is analyzed, and part of the buffer resource is reserved for the competition, the critical path resource requirement, the competition resource requirement of the non-critical path and the critical path overlap period and the buffer resource are added to predict the first computing resource that can guarantee the smooth execution of the entire task sequence with a high probability. By identifying the critical path and analyzing the resource requirements, the smoothness of the core link can be ensured, the system execution efficiency can be improved, by considering the resource competition of the non-critical path, potential resource conflict bottlenecks in the system can be predicted, the total resource amount predicted has the ability to cope with resource contention in actual execution, and the accuracy and practicability of resource prediction are improved, which provides data support for the formulation of overall resource allocation strategy.

[0089] Further, a dynamic subgraph matching mechanism is configured to map the operation intention vector to the pre-constructed task graph to identify task roles and obtain a task sequence, including:

[0090] S401, identifying conflict points of different terminal operation intentions at the same time sequence position according to the operation intention vector, performing intention deduplication or intention fusion at the conflict points to obtain a first intention vector;

[0091] S402, analyzing the distance between the operation entity and the task node in the pre-constructed task graph based on the first intention vector, calculating the intention response degree of each task node, and screening out task nodes with an intention response degree greater than a preset response threshold as a first task node set;

[0092] S403, analyzing the structural similarity of the operation intention and the task node execution template in the first task node set to identify task deviations;

[0093] S404, calculating the agent weight corresponding to each task node in combination with the task deviation and different task roles, assigning the task to the corresponding agent, and obtaining a task sequence.

[0094] In the embodiment, the conflict points of different terminal operation intentions at the same time sequence position are identified according to the operation intention vector, the cosine similarity of the operation intention vector in the same time sequence window is calculated, the cosine similarity threshold is set according to the intention analysis precision requirement, and the positions with the cosine similarity greater than the cosine similarity threshold are extracted as the conflict points; the logical association of the operation intentions in the conflict points is analyzed by using the convolutional neural network pre-trained by a large amount of historical data, the operation intentions without logical association are processed for deduplication, and the first operation intention is retained; the operation intentions with logical association are processed for mean fusion; and the deduplicated or fused intention vector is taken as the first intention vector. Through the conflict identification and deduplication or fusion processing, the individual intentions contradictory to each other are integrated into independent intentions, which provides a data basis for generating a task sequence and improves the coordination and operation efficiency of multi-agent cooperation.

[0095] Specifically, the distance between the operation entity and the task node in the pre-constructed task graph is analyzed based on the first intention vector, the operation entity node corresponding to the first intention vector is located in the pre-constructed task graph, the graph distance between each task node and the operation entity node is calculated by counting the number of connected edges, the cosine similarity between the first intention vector and the feature vector of each task node is calculated, the reciprocal of the graph distance and the cosine similarity are averaged and fused to obtain the intention response degree of the task node to the current intention, all task nodes are sorted according to the intention response degree, and the nodes with the intention response degree exceeding a preset response threshold are screened out to obtain a first task node set. The intention response degree is calculated by comprehensively considering the semantic correlation and the graph structure proximity, which can effectively exclude a large number of irrelevant task nodes, reduce the search range of fine matching, reduce the calculation complexity, improve the system response speed, ensure that the candidate node set is highly related to the user intention in semantics and logic, and improve the quality of the generated task sequence.

[0096] Specifically, in the first task node set, a standard execution template associated with each task node in the first task node set is obtained from the database, the standard execution template defines a standard step sequence, sub-operations and operation order for completing a task, the operation intention is mapped into a corresponding operation structure diagram according to the standard execution template format, the structural similarity between the operation structure diagram of the user and the standard execution template is analyzed by a graph matching algorithm, the positions with the structural similarity less than a structural similarity threshold are taken as task deviations according to the system task analysis precision requirement; through the structural similarity analysis, the matched task nodes can be quickly screened out, the deviation positions on the operation process are located, accurate data support is provided for task allocation and resource scheduling, and the adaptability and robustness of the system to the environment are improved.

[0097] Specifically, the weights of the agents corresponding to each task node are calculated in combination with the task deviation and different task roles. For each task node, the task role and the corresponding agent type are analyzed, and the task deviation and the adaptation degree of the task role to the inherent ability of the agent are summed and averaged to fuse. The weight of each task node for each agent is calculated. The Hungarian optimization algorithm is used to analyze the agent weight. The tasks are assigned to the most suitable agent according to the weight, and the tasks are arranged in the order of task dependency and logic, to obtain a task sequence. Through the analysis of the task deviation for the allocation decision, the dynamic adaptability and fault tolerance of the task allocation strategy can be improved. In combination with the weight calculation of the role ability, it is ensured that each task can be assigned to the agent that can most efficiently and reliably complete the task. The task sequence not only accurately reflects the user's intention, but also can improve the execution efficiency and task success rate of the multi-agent collaboration process in combination with the execution efficiency, resource optimization and abnormal processing ability.

[0098] Further, the coupling relationship between each agent and the tasks in different task roles is set, the value of the corresponding agent combination executing the task sequence is calculated, and the first computing resource is allocated to the corresponding agent according to the value to obtain a first resource allocation strategy, comprising:

[0099] S501, analyze the adaptation degree of different task roles in the task sequence to the corresponding task node agent, construct a coupling relationship, analyze the resource change of each task node in the terminal switching process according to the coupling relationship, and calculate the task execution priority;

[0100] S502, according to the task execution priority from high to low, simulate the value of different agent combinations executing the task sequence, and allocate the first computing resource to the corresponding agent according to the value to obtain a first resource allocation strategy.

[0101] In this embodiment, the adaptation degree of different task roles in the task sequence to the corresponding task node agent is analyzed, and a coupling relationship is constructed. For each task node in the task sequence, the required ability profile of the task role to which it belongs is analyzed, including but not limited to high real-time performance, strong graphics rendering capability, accurate logical judgment, and matching degree calculation with the inherent ability attributes of the candidate agent. The adaptation degree of the agent to the task node is calculated by a multi-factor weighted evaluation function to form a coupling relationship matrix. Based on the coupling relationship matrix, the resource change caused by the differences in terminal computing power, network delay and rendering load in the terminal switching process is analyzed. The influence of the switching process on the resource demand of the associated agent is simulated. The sum and average fusion calculation of the task's inherent criticality, coupling relationship strength and resource fluctuation risk caused by terminal switching is performed to calculate the task execution priority of each task node.

[0102] It should be noted that, by coupling the relationship based on the ability matching analysis, the priority evaluation analysis process not only depends on the static task topology, but also combines the characteristics of the execution subject, ensures that the high-priority task can be assigned to the most skilled agent for execution, and through the analysis of the resource change in the terminal switching process, the priority calculation can predictively respond to the dynamic fluctuations of the system environment, so that the priority of the key task with fierce resource competition or susceptible to environmental changes can be improved in advance, and the resource reservation is provided for the accurate sorting basis of the resource allocation process, and the task scheduling order is optimized.

[0103] Specifically, according to the task execution priority from high to low, the value of simulating different agent combination executing task sequence is obtained, the first computing resource is allocated to the corresponding agent according to the value, and the first resource allocation strategy is obtained. By simulating the execution effect of different agent combinations, the advantages and disadvantages of various allocation schemes can be comprehensively evaluated, local optimal decision is avoided, accurate data support is provided for the optimization scheduling of the multi-agent system, and the efficiency and effectiveness of the multi-agent cooperation process are improved.

[0104] As shown in Figure 3 According to the task execution priority from high to low, the value of simulating different agent combination executing task sequence is obtained, the first computing resource is allocated to the corresponding agent according to the value, and the first resource allocation strategy is obtained, including:

[0105] S601, according to the task execution priority from high to low, the concurrent execution process of different agent combinations in each task node is simulated, the resource consumption and teaching effect in the execution process are analyzed, and the value of the execution process is calculated.

[0106] S602, based on the value, the marginal benefit of resource demand in the first computing resource to different agents is analyzed, and the resource mapping relationship of the agent is constructed combined with the resource demand of each agent under different terminals.

[0107] S603, according to the agent resource mapping relationship, the first computing resource is allocated to the corresponding agent, and the first resource allocation strategy is obtained.

[0108] In the embodiment, a discrete event simulation model is constructed according to the task execution priority from high to low, the discrete event simulation model can simulate the concurrent execution behavior of multiple agents in a distributed environment, according to the order of task execution priority, considering the dependency relationship and timing constraint between tasks, in the process of simulating different agent combination executing task sequence in the model, real-time collection of resource consumption data of each task, including but not limited to CPU occupancy, memory usage, network bandwidth occupancy and storage IO throughput, analysis of task completion, operation accuracy, step coherence, response timeliness and other dimensions, through a multi-objective value function, the resource consumption index and the teaching effect index are normalized and weighted fusion, the resource consumption index is counted with negative weight, and the teaching effect index is counted with positive weight, the execution process value of each agent combination scheme is calculated.

[0109] It should be noted that through concurrent execution simulation, the resource competition and performance bottleneck in the cooperation of multiple agents can be accurately captured, and the resource consumption and teaching effect are analyzed, so that the resource allocation scheme can meet the efficiency requirement and guarantee the training quality, which can avoid the limitation of considering only a single performance index in the traditional resource allocation, provide comprehensive and accurate decision basis for resource optimization, and improve the scientificity and effectiveness of resource allocation.

[0110] Specifically, based on the value, the marginal benefit of resource demand in the first computing resource to different agents is analyzed, the value improvement effect of incremental allocation of each type of resource unit in the first computing resource to each agent is tested, for each agent, the response curve of resource input and value output is analyzed, the slope of the curve at each point is calculated as the marginal benefit value, the resource demand of each agent under different terminals is analyzed, including but not limited to the influence degree of terminal characteristics on agent resource demand, combining the marginal benefit analysis result and the terminal adaptation analysis result, the agent resource mapping relationship is constructed through a multi-constraint optimization algorithm, the agent resource mapping relationship determines the expected benefit that each type of resource allocated to a specific agent can produce in a given terminal environment. Through quantitative analysis of the improvement effect of resource increment on agent efficiency, the benefit critical point of resource input can be accurately identified, combined with resource demand analysis under different terminal environments, the agent resource mapping relationship has environment self-adaptation ability, which can ensure that limited computing resources always flow to the agent with the highest use efficiency, effectively avoiding resource waste and diminishing returns, and improving the effectiveness of resource allocation strategy.

[0111] Specifically, according to the agent resource mapping relationship, the resource units are allocated to the corresponding agents in descending order of marginal benefit; in the allocation process, the lowest resource guarantee requirement of each agent, the unit granularity of resource allocation, the load balancing degree between agents and the fault tolerance requirement of the system are combined, and the iterative calculation is performed through the priority-based greedy algorithm, so that the overall value is maximized under the premise of meeting the constraint conditions, and a first resource allocation strategy is obtained. Through the resource allocation mechanism based on priority and constraint conditions, the resource supply of high-value tasks is guaranteed, the stability of system operation is improved, the overall operation efficiency and resource utilization of the system are improved while meeting the multi-agent collaboration demand, and reliable resource guarantee is provided for stable operation of the complex training environment.

[0112] Further, according to the first resource allocation strategy, a task execution process of the agent combination executing the task sequence is simulated, a fault link in the task execution process is identified by adding a simulated disturbance, the first resource allocation strategy is optimized, and a second resource allocation strategy is obtained, including:

[0113] S701, according to the first resource allocation strategy, a task execution process of the agent combination executing the task sequence is simulated, a fault link in the task execution process is identified by adding a simulated disturbance;

[0114] S702, according to the fault link, the correlation between different fault types and resource allocation is analyzed, the first resource allocation strategy is optimized, and a second resource allocation strategy is obtained.

[0115] In this embodiment, according to the first resource allocation strategy, a task execution process of the agent combination executing the task sequence is simulated, a fault link in the task execution process is identified by adding a simulated disturbance; by actively injecting a simulated disturbance, the potential defects and weak links of the resource allocation scheme under various abnormal conditions can be exposed without affecting the actual production environment, the fault link and its occurrence conditions can be accurately identified, a clear target and direction are provided for strategy optimization, the fault tolerance and robustness of the system are enhanced, the system performance degradation or service interruption caused by unexpected changes in the operating environment can be effectively prevented, and the reliability and stability of the system operation process are improved.

[0116] Specifically, according to the analysis of the association between different fault types and resource allocation, the identified fault links are classified and root cause analyzed, and the fault types include but are not limited to resource insufficient type, resource competition type, and resource allocation uneven type. The association between different fault types and resource allocation is analyzed, including analyzing the statistical relationship between factors such as resource allocation intensity, resource allocation timing, resource combination configuration and fault occurrence rate when the fault occurs. The fault risk of a specific resource allocation scheme under different interference scenarios is quantified. The overall fault risk is reduced as the target, and the first resource allocation strategy is iteratively adjusted through a multi-objective optimization algorithm. The optimization includes but is not limited to adding resource redundancy for critical tasks, adjusting resource allocation timing to avoid competition, rebalancing the load allocation between agents, setting up a dynamic resource recycling and reallocation mechanism. After iterative optimization, the second resource allocation strategy is obtained.

[0117] It should be noted that the optimization of the resource allocation strategy can improve the pertinence and effectiveness of the strategy adjustment. Through resource redundancy and dynamic adjustment mechanism, the environmental adaptability and fault tolerance of the resource allocation strategy are improved, which can effectively reduce the fault occurrence rate of the system in actual operation, improve the continuity and reliability of the service, and at the same time, through optimizing the resource use efficiency, the resource waste caused by excessive conservative resource allocation is avoided.

[0118] Further, according to the first resource allocation strategy, a task execution process of the agent combination executing the task sequence is simulated, and a fault link in the task execution process is identified by adding simulated interference, including:

[0119] S801, simulating a task execution process of an agent combination executing the task sequence according to the first resource allocation strategy, and obtaining a first execution result;

[0120] S802, based on the first execution result, extracting the interference mode corresponding to the multi-terminal training process, calculating the adaptation degree of each interference mode to the training process, and constructing an interference matrix;

[0121] S803, according to the interference matrix, adding corresponding interference in the task execution simulation process, analyzing the fault condition, and identifying the fault link.

[0122] In this embodiment, the agent combination performs the task execution process according to the first resource allocation strategy, and the agent combination performs the complete task sequence in the simulation environment according to the resource quota and scheduling timing determined according to the first resource allocation strategy. During the simulation execution process, the system continuously collects the running state data of each agent, including but not limited to task start and end time, resource utilization curve, intermediate product generation state, and communication record between agents. Key performance indicators are recorded, including but not limited to total task sequence completion time, actual time consumption of each task node, resource utilization efficiency, etc. The key performance indicators are cleaned and aggregated to obtain the first execution result. By obtaining the complete first execution result, the performance baseline in the ideal state can be obtained, which provides a data basis for identifying performance deviation and system vulnerability, can comprehensively evaluate the effectiveness of the resource allocation strategy under ideal conditions, provides a basis for comparison for interference impact analysis, and ensures the accuracy and pertinence of the optimization direction.

[0123] Specifically, based on the first execution result, the task execution characteristics and resource usage mode in the first execution result are analyzed, the key dependency points and performance sensitive links of the system are identified, and interference modes matched with the characteristics of the multi-terminal training process are extracted from the interference mode library, including but not limited to network bandwidth fluctuation, terminal device performance degradation, data transmission packet loss, and computing resource burst competition. For each interference mode, the association strength of the interference mode with the terminal type, the matching degree with the task criticality, and the degree of agreement with the historical fault data are analyzed, the correlation degree of the interference mode with the current training process is calculated, an interference matrix is constructed, the rows of the interference matrix represent different interference types, the columns represent the affected task nodes or agents, and the matrix element values represent the adaptation weight of a specific interference to a specific target. Through adaptation degree calculation, the interference injection is more concentrated in the system vulnerable links, avoiding the blindness of testing and resource waste, and through the construction of the interference matrix, accurate data support is provided for the system interference analysis process, improving the efficiency and accuracy of fault identification.

[0124] Specifically, based on the interference type, target node and injection intensity determined by the interference matrix, the corresponding interference is dynamically added in the task execution simulation process, the response state of the system is monitored in real time, and the fault condition is analyzed, including but not limited to performance indicators exceeding the standard, business process interruption, and resource exhaustion alarm. When abnormal phenomena are monitored, the time point of fault occurrence, the range of influence, the recovery time and other detailed information are immediately recorded, and the fault link that causes the root cause of the fault is identified, including but not limited to resource bottleneck of a specific agent and resource competition conflict between tasks. By adding interference, the performance bottleneck of the system under stress conditions can be efficiently and accurately exposed, and through fault identification analysis, the corresponding improvement direction is provided for strategy optimization, the robustness and fault tolerance capability of the system are improved, potential problems are found and repaired before actual deployment, and the risk of faults in the production environment is effectively reduced.

[0125] Further, according to the second resource allocation strategy, each agent is allocated corresponding computing resources, and each agent is scheduled to execute each task in the task sequence to optimize the multi-agent collaboration process, including:

[0126] S901, according to the second resource allocation strategy, each agent is allocated corresponding computing resources;

[0127] S902, in combination with the computing resource allocation and the task execution priority, each agent is scheduled to execute each task in the task sequence to optimize the multi-agent collaboration process.

[0128] In this embodiment, according to the second resource allocation strategy, each agent is allocated corresponding computing resources; the resource allocation details defined in the second resource allocation strategy are analyzed, including but not limited to the type, quantity, quality attribute and allocation duration of the computing resources that each agent should obtain, and the resources are allocated; in the allocation process, resource isolation technology is used to ensure that the computing resources obtained by each agent do not interfere with each other, and the actual resource usage of each agent is tracked in real time. Based on the strategy allocation resource, the accurate matching of resource supply and task demand can be ensured, and the performance bottleneck caused by insufficient resources or the waste caused by excessive resources can be avoided, and through resource isolation, the stability and reliability of the multi-agent parallel execution can be ensured.

[0129] Specifically, in combination with the computing resource allocation and the task execution priority, each agent is scheduled to execute each task in the task sequence to ensure that high-priority tasks can be triggered first, and the agents allocated to them are reserved sufficient start-up and running resources, the task groups that can be executed concurrently are identified, and the idle resources available in the system are used to make them parallel to shorten the overall completion time; the task execution status and resource consumption of each agent are continuously monitored, and once a task execution delay or resource anomaly of an agent is found, the preset fault tolerance strategy is used for timely intervention, including but not limited to rescheduling tasks or adjusting resource allocation, to ensure that the entire task sequence can be executed smoothly and efficiently, and finally the systematic optimization of the multi-agent collaboration process is realized.

[0130] As shown in Figure 4 the multi-agent collaboration optimization system based on multi-terminal training access is used to realize the multi-agent collaboration optimization method based on multi-terminal training access, including:

[0131] An operation intention analysis module extracts operation actions, operation entities and corresponding constraint conditions through context association analysis according to multi-modal training operation data from different terminals, and constructs an operation intention vector;

[0132] a computing resource prediction module, which maps the operation intention vector to a pre-constructed task graph, identifies a task sequence through role decomposition, analyzes historical performance data of corresponding tasks in the task sequence, and predicts first computing resources;

[0133] a resource allocation module, which sets a coupling relationship between each agent and a task in different task roles, calculates a value of executing the task sequence by a corresponding agent combination, allocates the first computing resources to the corresponding agent according to the value, and obtains a first resource allocation strategy;

[0134] a resource optimization module, which simulates a task execution process of the agent combination executing the task sequence according to the first resource allocation strategy, identifies a fault link in the task execution process by adding a simulated disturbance, optimizes the first resource allocation strategy, and obtains a second resource allocation strategy;

[0135] a multi-agent collaboration module, which allocates corresponding computing resources to each agent according to the second resource allocation strategy, schedules each agent to execute each task in the task sequence, and optimizes a multi-agent collaboration process.

[0136] In this embodiment, the operation intention analysis module receives and analyzes multi-modal training operation data from different terminals, extracts key semantic elements such as operation actions, operation entities and corresponding constraint conditions from the data through context association analysis, constructs an operation intention vector that can reflect user intention, and extracts corresponding data through analysis, which can improve the understanding ability of the system for the training scene and effectively avoid analysis errors caused by misjudgment of intention; the computing resource prediction module maps the operation intention vector to a pre-constructed task graph through a dynamic subgraph matching mechanism, decomposes a task sequence with role attribution, predicts first computing resources that meet the computing demand based on historical performance data and dependency relationships of tasks in the task sequence, combines task planning and resource prediction, provides accurate and reliable data support for pre-allocation of required resources for the system, and improves resource utilization.

[0137] Specifically, the resource allocation module establishes a coupling relationship model between the agent and the task role, calculates the comprehensive value by simulating the process of different agent combinations executing the task sequence, allocates the first computing resource to the corresponding agent according to the value maximization principle, obtains the first resource allocation strategy, and ensures that the limited computing resource is preferentially invested in the agent that can produce the maximum overall efficiency through value allocation, thereby improving the marginal benefit of the computing resource and the collaboration efficiency of the system; the resource optimization module optimizes the first resource allocation strategy by simulating the process of the agent combination executing the task and adding a fault identification link of simulated interference, obtains the second resource allocation strategy, improves the fault tolerance and anti-interference ability of the system, exposes and repairs the strategy weakness through fault simulation, improves the reliability of the resource strategy in response to the complex environment, and improves the stability of the system in actual operation; the multi-agent collaboration module allocates the corresponding computing resource to each agent according to the optimized second resource allocation strategy, and schedules the agent to execute the corresponding task according to the task sequence and priority, so that multiple agents work collaboratively, the optimization of the multi-agent collaboration process is realized, and the system operation performance is improved.

[0138] The above only describes the preferred embodiments of the present application, and the protection scope of the present application is not limited to the above-described embodiments. Any technical solution falling within the scope of the present application should be considered as falling within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application should also be considered as falling within the protection scope of the present application.

Claims

1. A multi-agent collaborative optimization method based on multi-terminal training access, characterized in that, The method comprises the following steps: According to the multi-modal training operation data from different terminals, analyze the pause of operation action and the switching rule of teaching link, identify the operation process boundary point, divide the multi-modal training operation data into multiple operation segments, and obtain an operation segment set; Based on the operation segment set, perform feature extraction through a feature extraction model to obtain a feature sequence; According to the feature sequence, analyze the interaction state and operation tool usage of the operation process, construct the mapping relationship between the operation entity and the training knowledge, and calculate the operation entity correlation matrix; In the operation entity correlation matrix, analyze the abnormal behavior mode and the operation habit of different users, extract the time constraint and the abnormal condition, and construct a constraint condition set; Combine the operation segment set, the operation entity correlation matrix and the constraint condition set, analyze the teaching structure, the operation correlation and the constraint priority, and construct an operation intention vector; Map the operation intention vector into a pre-constructed task graph, identify the task sequence by decomposing the task role, analyze the historical performance data of the corresponding task in the task sequence, and predict a first computing resource; Set the coupling relationship between each agent and the task in different task roles, calculate the value of the corresponding agent combination executing the task sequence, allocate the first computing resource to the corresponding agent according to the value, and obtain a first resource allocation strategy; According to the first resource allocation strategy, simulate the task execution process of the agent combination executing the task sequence, identify the fault link in the task execution process by adding simulation interference, optimize the first resource allocation strategy, and obtain a second resource allocation strategy; According to the second resource allocation strategy, allocate the corresponding computing resource to each agent, and schedule each agent to execute each task in the task sequence to optimize the multi-agent collaboration process.

2. The multi-agent collaborative optimization method based on multi-terminal training access according to claim 1, characterized in that, Mapping the operation intention vector into a pre-constructed task graph, identifying the task sequence by decomposing the task role, analyzing the historical performance data of the corresponding task in the task sequence, and predicting a first computing resource, comprises: Configure a dynamic subgraph matching mechanism, map the operation intention vector into a pre-constructed task graph, identify the task sequence by decomposing the task role, and extract the task features according to the task sequence. According to the task sequence, analyze the historical performance data of each task, extract the task features, and calculate the first performance data of each task. Analyze the dependency relationship between tasks, identify the critical path in the task sequence according to the dependency relationship, combine the task resource demand situation of the critical path and the resource overlapping competition situation of the non-critical path, and predict the first computing resource that meets the computing demand of the task sequence.

3. The multi-agent collaborative optimization method based on multi-terminal training access according to claim 2, characterized in that, The configuration of the dynamic subgraph matching mechanism, the mapping of the operation intention vector into a pre-constructed task graph, the identification of the task sequence by decomposing the task role, and the extraction of the task features according to the task sequence, comprise: According to the operation intention vector, identify the conflict points of different terminal operation intentions at the same time sequence position, perform intention deduplication or intention fusion at the conflict points, and obtain a first intention vector; Based on the first intention vector analysis operation entity and the distance between the task nodes in the pre-constructed task graph, the intention response degree of each task node is calculated, and the task nodes with an intention response degree greater than a preset response threshold are selected as a first task node set; In the first task node set, the structural similarity between the operation intention and the task node execution template is analyzed, and a task deviation is identified; In combination with the task deviation and different task roles, the weight of each task node corresponding to the intelligent agent is calculated, the task is assigned to the corresponding intelligent agent, and a task sequence is obtained.

4. The multi-agent collaborative optimization method based on multi-terminal training access according to claim 1, characterized in that, The coupling relationship between each intelligent agent and the tasks in different task roles is set, the value of the corresponding intelligent agent combination executing the task sequence is calculated, and the first computing resource is allocated to the corresponding intelligent agent according to the value to obtain a first resource allocation strategy, including: The adaptation degree of different task roles and corresponding task node intelligent agents in the task sequence is analyzed, a coupling relationship is constructed, the resource change of each task node in the terminal switching process is analyzed according to the coupling relationship, and the task execution priority is calculated; According to the task execution priority from high to low, the value of different intelligent agent combinations executing the task sequence is simulated, and the first computing resource is allocated to the corresponding intelligent agent according to the value to obtain a first resource allocation strategy.

5. The multi-agent collaborative optimization method based on multi-terminal training access according to claim 4, characterized in that, According to the task execution priority from high to low, the value of different intelligent agent combinations executing the task sequence is simulated, and the first computing resource is allocated to the corresponding intelligent agent according to the value to obtain a first resource allocation strategy, including: According to the task execution priority from high to low, the concurrent execution process of different intelligent agent combinations in each task node is simulated, the resource consumption and teaching effect in the execution process are analyzed, and the value of the execution process is calculated; Based on the value, the marginal benefit of resource demand in the first computing resource to different intelligent agents is analyzed, and the intelligent agent resource mapping relationship is constructed in combination with the resource demand of each intelligent agent under different terminals; According to the intelligent agent resource mapping relationship, the first computing resource is allocated to the corresponding intelligent agent to obtain a first resource allocation strategy.

6. The method of claim 1, wherein, According to the first resource allocation strategy, the task execution process of the intelligent agent combination executing the task sequence is simulated, the fault link in the task execution process is identified by adding a simulated disturbance, and the first resource allocation strategy is optimized to obtain a second resource allocation strategy, including: According to the first resource allocation strategy, the task execution process of the intelligent agent combination executing the task sequence is simulated, the fault link in the task execution process is identified by adding a simulated disturbance; According to the fault link, the association between different fault types and resource allocation is analyzed, and the first resource allocation strategy is optimized to obtain a second resource allocation strategy.

7. The multi-agent collaboration optimization method based on multi-terminal training access according to claim 6, characterized in that, According to the first resource allocation strategy, the task execution process of the intelligent agent combination executing the task sequence is simulated, the fault link in the task execution process is identified by adding a simulated disturbance, including: According to the first resource allocation strategy, the task execution process of the intelligent agent combination executing the task sequence is simulated, and a first execution result is obtained; Based on the first execution result, the interference mode corresponding to the multi-terminal training process is extracted, the adaptation degree of each interference mode to the training process is calculated, and an interference matrix is constructed; According to the interference matrix, corresponding interference is added in a task execution simulation process, a fault condition is analyzed, and a fault link is identified.

8. The method of claim 1, wherein, According to the second resource allocation strategy, each agent is allocated corresponding computing resources, and each agent is scheduled to execute each task in the task sequence, so as to optimize the multi-agent collaboration process, including: According to the second resource allocation strategy, each agent is allocated corresponding computing resources; In combination with the computing resource allocation and the task execution priority, each agent is scheduled to execute each task in the task sequence, so as to optimize the multi-agent collaboration process.

9. A multi-agent collaborative optimization system based on multi-terminal training access, characterized in that, The method comprises the following steps: An operation intention analysis module extracts operation actions, operation entities and corresponding constraint conditions by context association analysis according to multi-modal training operation data from different terminals, and constructs an operation intention vector; A computing resource prediction module maps the operation intention vector into a pre-constructed task graph, identifies task role decomposition to obtain a task sequence, analyzes historical performance data of corresponding tasks in the task sequence, and predicts a first computing resource; A resource allocation module sets a coupling relationship between each agent and a task in different task roles, calculates a value of a combination of corresponding agents executing the task sequence, allocates the first computing resource to the corresponding agents according to the value, and obtains a first resource allocation strategy; A resource optimization module simulates a task execution process of the combination of agents executing the task sequence according to the first resource allocation strategy, identifies a fault link in the task execution process by adding simulation interference, optimizes the first resource allocation strategy, and obtains a second resource allocation strategy; A multi-agent collaboration module allocates corresponding computing resources to each agent according to the second resource allocation strategy, and schedules each agent to execute each task in the task sequence, so as to optimize the multi-agent collaboration process.

Images