Algorithm network intelligent management method, computer device and storage medium

Abstract

This application relates to the field of artificial intelligence technology and discloses a computing network intelligent management method, computer device, and storage medium. The method includes: determining a task type in response to a current task instruction, and determining a first type of intelligent agent, a second type of intelligent agent, and a third type of intelligent agent matching the task type; generating multiple current sub-tasks using the first type of intelligent agent; executing each current sub-task using the second type of intelligent agent and outputting the execution result; and, in the case of an operation-type task, dynamically adjusting the execution of the current sub-task based on the execution status during its execution using the third type of intelligent agent, or triggering the first type of intelligent agent to dynamically adjust the computing network management strategy based on the execution status after the current sub-task has finished executing. This application enables computing network scheduling to efficiently meet user needs, thereby improving the execution efficiency and reliability of computing network scheduling tasks.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a computer network intelligent management method, computer equipment and storage medium. Background Technology

[0002] Computing power scheduling and management is a crucial component of technologies such as artificial intelligence, big data, and cloud computing. Most computing network scheduling and management platforms rely on pre-defined, fixed processes and rules for resource management and task scheduling, resulting in poor flexibility and difficulty in adapting to dynamic changes in the computing network execution environment. With the development of large-scale modeling technology, intelligent agents have begun to be applied to computing network scheduling and management platforms.

[0003] However, in related technologies, intelligent agents usually only focus on their own task execution. Due to the complexity and variability of the execution environment of the computing network scheduling, the execution results output by the intelligent agent may not meet the user's needs, which seriously affects the execution efficiency and reliability. Summary of the Invention

[0004] This application provides a computing network intelligent management method, computer equipment, and storage medium, which solves the technical problem that current computing network scheduling and management platforms are unable to adapt to complex and ever-changing execution environments, affecting execution efficiency and reliability. By constructing a multi-agent collaborative architecture with role-based division of labor, dynamic iteration and adjustment of tasks are realized to enhance the system's adaptability to the execution environment, enabling computing network scheduling to efficiently meet user needs, thereby improving the execution efficiency and reliability of computing network scheduling tasks.

[0005] To achieve the above objectives, the main technical solutions adopted in this application include: Firstly, this application provides a method for intelligent management of a computing network, the method comprising: In response to the current task instruction, a task type is determined, and a task group matching the task type is determined, the task group including a first type of intelligent agent, a second type of intelligent agent, and a third type of intelligent agent; The first type of intelligent agent generates a computing network management strategy based on the current task instruction, and the computing network management strategy includes multiple current sub-tasks; The second type of intelligent agent is used to execute each of the current sub-tasks and output the execution results of the current sub-tasks; When the task type is an operation task, the second type of intelligent agent sends the execution status of the current sub-task to the third type of intelligent agent, so that the third type of intelligent agent can adjust the current sub-task according to the execution status during the execution of the current sub-task, or trigger the first type of intelligent agent to dynamically adjust the computing network management strategy according to the execution status after the current sub-task is completed.

[0006] The proposed intelligent network management method identifies matching task groups for different task types, and each task group includes multiple agents with different roles. First-type agents generate network management strategies based on current task instructions, achieving precise task instruction breakdown and strategy formulation. Second-type agents execute the various sub-tasks within the network management strategy. Furthermore, if the task type is an operational task, a third-type agent coordinates the execution of each sub-task. This includes adjusting the execution process of the second-type agents or reporting the adjustments to the first-type agents to generate a more suitable network management strategy for the actual network environment. This ensures information synchronization and task collaboration among the agents, resulting in network management outcomes that meet user needs. Compared to related technologies, this application, by constructing a multi-agent collaborative architecture with role-based division of labor, achieves dynamic iteration and adjustment of task plans, significantly enhancing adaptability to the dynamic environment of network scheduling, thereby improving the execution efficiency and reliability of network scheduling tasks.

[0007] Secondly, this application provides a computer device, comprising: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes these computer instructions to perform the aforementioned intelligent management method for computer networks.

[0008] Thirdly, this application provides a computer-readable storage medium storing computer instructions, which are used to cause a computer to execute the above-described intelligent management method for computer networks. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0010] Figure 1 This is one of the flowcharts illustrating a computer network intelligent management method proposed in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a computing network intelligent management device proposed in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device proposed in an embodiment of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] Computing power scheduling refers to the process of dynamically allocating computing resources in a distributed, multi-node computing environment based on factors such as task priority, resource requirements, and real-time load to achieve optimal system performance and resource utilization efficiency. The core of computing power scheduling technology lies in using intelligent algorithms to efficiently allocate computing resources to tasks that need processing, avoiding resource idleness or overload. It is widely used in fields such as cloud computing, big data processing, edge computing, and artificial intelligence model training.

[0013] In some application scenarios, network scheduling and management platforms mostly rely on preset fixed processes and rules for resource management and task scheduling, resulting in poor flexibility and difficulty in adapting to dynamically changing business needs. Users often need to search through multiple management interfaces or rely on specialized query languages ​​to view specific metrics, leading to high operational barriers and low efficiency. Furthermore, current network scheduling and management platforms overly rely on human experience when handling complex operation and maintenance tasks, with insufficient automation and intelligence, resulting in slow fault response and high error rates.

[0014] To address this, some related technologies have introduced a single agent into the computing network scheduling and management platform. However, configuring too many tools for a single agent can lead to a decline in its tool selection effectiveness in complex tasks. Furthermore, when handling long-process tasks, the message context grows rapidly, resulting in a surge in token consumption and high costs. In addition, the lack of division of labor and collaboration within a single agent makes it difficult to handle comprehensive tasks in computing network scheduling that involve knowledge from multiple domains (such as resources, network, and security).

[0015] Multi-agent systems based on large models commonly employ a plan-and-execute (PAG) model. In this model, a planning agent pre-defines a fixed plan, which is then executed by the executing agents in a cyclical manner, following a process of reasoning, action, and re-act. However, this model has significant drawbacks: First, the computational network execution environment is highly dynamic, frequently encountering unpredictable problems such as network interruptions and tool call failures. Once the plan is finalized, the planning agent no longer participates in subsequent execution, failing to perceive environmental changes and adjust accordingly, making the task prone to failure when environmental anomalies occur. Second, each executing agent typically focuses only on its own subtask, lacking a global understanding of the overall task progress, leading to duplicated work or operational conflicts, impacting execution efficiency and reliability.

[0016] Therefore, there is an urgent need for an intelligent, adaptive, and highly reliable computing network scheduling and management method.

[0017] According to an embodiment of this application, a method for intelligent management of a computing network is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0018] This embodiment provides a computing network intelligent management method, applied to a computing network scheduling and management platform. Figure 1 This is a flowchart of a computer network intelligent management method according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps: Step S1: In response to the current task instruction, determine the task type and determine the task group that matches the task type. The task group includes a first type of intelligent agent, a second type of intelligent agent, and a third type of intelligent agent.

[0019] In this embodiment, the computing network scheduling and management platform is managed by a master agent. Specifically, the master agent receives and understands user requests and performs intent recognition to determine whether the user is simply engaging in a casual chat or requires the computing network scheduling and management platform to handle specific computing network management tasks. If it is determined to be a casual chat, the platform can directly invoke the capabilities of the large model to process the response. If the platform needs to handle specific computing network management tasks, the user request is treated as the current task instruction, and the corresponding task type is determined. For example, task types include, but are not limited to, operation tasks and query tasks.

[0020] It is understandable that different types of tasks have different processing flows and execution complexities. From the perspective of network resources, this application embodiment can be considered a query task if it does not change the resource status, such as "query which clusters a certain application is currently deployed in". Operations that require changing the resource status can be considered operation tasks, such as "open a virtual machine with 4 cores and 8GB of memory". Compared with query tasks, operation tasks usually involve changes in resource status.

[0021] In this embodiment, different task teams are created by the Master Agent for different task types. For example, for query tasks, task teams are composed of agents suitable for query tasks, and the agents in each task team have different roles, mimicking the collaborative model of a human team. The first type of agent is used for task parsing and network management strategy formulation, acting as the leader of the task team. The second type of agent is used for the specific execution of sub-tasks, acting as the executor. The third type of agent plays different roles depending on the task type. If the task type is an operation task, the third type of agent is used for global coordination and anomaly reporting during sub-task execution, acting as the coordinator. If the task type is a query task, the third type of agent is used to collect and integrate the execution results of the various second type of agents for the current sub-task to obtain the final query result, acting as the summarizer.

[0022] Step S3: The first type of intelligent agent generates a computing network management strategy based on the current task instructions. The computing network management strategy includes multiple current sub-tasks.

[0023] Specifically, the first type of intelligent agent first parses the current task instruction, and then, based on the parsing results, breaks down the current task instruction into multiple independently executable sub-tasks, thereby generating a computing network management strategy. In some embodiments of this application, during the parsing process of the current task instruction, the first type of intelligent agent accesses a preset knowledge base to obtain the information required for strategy generation, thereby ensuring the accurate breakdown of the current task instruction and improving the execution efficiency and accuracy of the computing network management strategy.

[0024] The pre-defined knowledge base is implemented based on RAG (Retrieval-Augmented Generation) technology. RAG technology combines information retrieval and text generation techniques, enhancing the generation capabilities of large models by retrieving relevant documents in real time. Furthermore, RAG technology helps address issues such as knowledge stagnation and illusions in large models for intelligent agents.

[0025] In some embodiments of this application, the information imported into the preset knowledge base includes: (1) Knowledge in the field of computing and networking, including general knowledge in professional fields such as cloud computing, containers, and networks; (2) Knowledge of operation and maintenance issues, including a knowledge manual on the causes of various problems and suggestions for solutions; (3) Application deployment knowledge, including application deployment preference (e.g., whether the application to be deployed is compute-intensive or I / O-intensive) modeling data, and related test benchmark data; (4) An expert guide to the task execution process, including the execution steps of the process and how to solve problems.

[0026] Step S5: Utilize the second type of intelligent agent to execute each current subtask and output the execution result of the current subtask.

[0027] Specifically, the second type of intelligent agent uses a pre-set knowledge base to execute the current sub-task. For operation-type tasks, the second type of intelligent agent will report information such as execution progress and execution results to the third type of intelligent agent in real time, so that the third type of intelligent agent can monitor the execution status of each current sub-task in real time.

[0028] Step S7: When the task type is an operation task, the second type of intelligent agent sends the execution status of the current subtask to the third type of intelligent agent in real time. The third type of intelligent agent can dynamically adjust the execution of the current subtask based on the execution status during the execution of the current subtask, or trigger the first type of intelligent agent to dynamically adjust the computing network management strategy based on the execution status after the current subtask is completed.

[0029] Specifically, for operational tasks, during the execution of the current subtask, the second type of intelligent agent actively reports the execution status to the third type of intelligent agent at each step of the current subtask. When the third type of intelligent agent receives the execution status of the current subtask, it checks for any execution conflicts or duplicates among the current subtasks. If any are found, it coordinates the relevant second type of intelligent agents to dynamically adjust the execution of the corresponding current subtask. On the other hand, if the second type of intelligent agent reports that the current subtask has failed, the third type of intelligent agent summarizes the failure information to generate an execution report and notifies the first type of intelligent agent, thereby returning to step S3 to realize the dynamic adjustment of the computing network management strategy.

[0030] The intelligent network management method provided in this application determines matching task groups for different task types, and each task group includes multiple intelligent agents with different roles. A first type of intelligent agent generates a network management strategy based on the current task instructions, achieving precise breakdown of task instructions and strategy formulation. A second type of intelligent agent executes each current sub-task in the network management strategy. Furthermore, if the task type is an operational task, a third type of intelligent agent coordinates the execution of each current sub-task. For example, it adjusts the execution process of the current sub-task by the second type of intelligent agent or reports it to the first type of intelligent agent to regenerate a network management strategy more suitable for the actual network environment. This ensures information synchronization and task collaboration among the intelligent agents, thereby obtaining network management results that meet user needs. Compared with related technologies, this application, by constructing a multi-agent collaborative architecture with role-based division of labor, achieves dynamic iteration and adjustment of task plans, significantly enhancing adaptability to the dynamic environment of network scheduling, thus improving the execution efficiency and reliability of network scheduling tasks.

[0031] It should be noted that the embodiments of this application will pre-configure a standardized toolset that can be understood and invoked by the large model of the intelligent agent, namely the MCP (Model Control Protocol) tool system. MCP defines a standard way for applications and large models to exchange context information, enabling developers to connect data sources, APIs, functional modules, etc. to the large model in a unified manner, solving the fragmentation problem of tool call interfaces of different models, and greatly reducing the complexity of intelligent agent development and integration.

[0032] Specifically, this application's embodiments abstract and encapsulate the core functions of the computing network scheduling and management platform, transforming them into semantically clear and interface-standardized MCP tools. Specifically, the platform's unified resource management capability is transformed into resource query and status awareness tools. By encapsulating access interfaces to computing nodes (cloud, edge, and endpoint) and network resources, the intelligent agent large model can obtain a real-time global resource view. The platform's intelligent scheduling and orchestration capabilities are toolified into scheduling algorithm recommendation interfaces, task scheduling, and resource operation interfaces, supporting the intelligent agent large model to dynamically initiate scheduling commands based on business needs, enabling tasks such as deployment, load balancing, and elastic scaling. The platform's real-time monitoring and analysis capabilities are transformed into data query and trend prediction tools. Monitoring indicators, log data, and prediction model outputs are set as callable interfaces, supporting the intelligent agent large model in perception and decision-making. The platform's automated operation and maintenance capabilities are encapsulated into operation and maintenance execution tools, such as fault recovery, configuration updates, backup and restart, enabling the intelligent agent large model to have autonomous operation and maintenance capabilities and automatically resolve problems. The platform's security mechanisms are tooled into interfaces for permission verification, policy application, and audit trails to ensure that the operation of the large-scale intelligent agent model complies with security and compliance requirements.

[0033] Through the aforementioned tool-based transformation, the various native capabilities of the computing network scheduling and management platform can be systematically transformed into a set of MCP tools that are understandable, schedulable, and executable for large models, laying the foundation for realizing agent-based computing network scheduling.

[0034] When the task type is a query task, step S3 above may include the following steps: Step S311: Use the first type of intelligent agent to parse the current task instruction to determine the first parsing result corresponding to the current task instruction.

[0035] Specifically, the current task instruction input by the user is text information. The first type of intelligent agent performs semantic parsing on the received current task instruction and obtains the parsing result. It can be understood that the parsing result represents the user's explicit query requirements.

[0036] Step S313: Determine multiple current subtasks in parallel based on the first parsing result.

[0037] Specifically, the first type of intelligent agent accesses a preset knowledge base based on the above query requirements, obtains relevant knowledge from the preset knowledge base, breaks down the query requirements into multiple parallel current subtasks, and then assigns these current subtasks to the second type of intelligent agent with corresponding query capabilities.

[0038] This application embodiment decomposes a task into multiple subtasks that can be processed in parallel by a first type of intelligent agent, so that these subtasks can be subsequently assigned to a second type of intelligent agent for parallel processing, thereby helping to solve the problem of massive data computation and improving the execution efficiency of query tasks.

[0039] Furthermore, in some embodiments of this application, MapReduce is a programming model for parallel computation of large-scale datasets (greater than 1TB). It solves the problem of massive data computation by decomposing complex tasks into Map and Reduce stages and processing data in parallel. Embodiments of this application implement the above steps based on the MapReduce method, decomposing complex query tasks into multiple parallel-executable subtasks, and then assigning each subtask to different second-type intelligent agents, which then execute their respective subtasks.

[0040] Based on the MapReduce method, for query-type tasks, the first type of agent is the query leader, the second type is the execution agent (Map Agent), and the third type is the integration agent (Reduce Agent). The following example, using the task instruction "to query the deployment location of a specific application," details the execution flow of query-type tasks: Step 1: The user enters a query command for the current task. Step 2: The Query Leader parses the current task instructions received from the user: (1) Decompose the current task instruction into a series of independent, parallel-executable current subtasks, such as querying which clouds the application is deployed on, querying the interfaces of multiple clouds, etc. (2) Assign different tasks according to the different tool capabilities of the Map Agent. For example, the task of querying public cloud information is assigned to the intelligent agent that has public cloud query tools. Step 3: Each Map Agent independently executes its assigned subtask, and multiple Map Agents process tasks in parallel. Step 4: The Reduce Agent collects the data returned by each Map Agent, and after all Map Agents have finished executing, it integrates the execution results of each current subtask and outputs them to the Master Agent.

[0041] In this embodiment, the Master Agent aggregates the execution results of all current subtasks and generates output conforming to user instructions or formats as the computing network management result to complete computing network scheduling.

[0042] In some embodiments of this application, for query-type tasks, since there are usually multiple users in the system, each with different data query permissions, and some data may contain user privacy or sensitive information that may be leaked during the execution of the large model agent, this application embodiment also identifies a fourth type of agent matching the query-type task, and the method further includes: After all current subtasks have been executed, the fourth type of intelligent agent is used to obtain the permission and risk information corresponding to the current task instruction. Specifically, the fourth type of intelligent agent is the Audit Agent. After receiving a user's query task instruction, the Audit Agent queries the current user's permission information to form a permission list containing the aforementioned permission information. Simultaneously, it retrieves relevant documents from a pre-set knowledge base. These documents contain rules regarding privacy risks that require attention, such as not disclosing user-deployed application access data, or containing private fields such as user passwords, thereby generating a risk list containing risk information. Understandably, for each query task request, the Audit Agent reconstructs the permission and risk lists to ensure timely and accurate verification.

[0043] The third type of intelligent agent is used to obtain the query results of each current subtask, and the query results are filtered according to permission information and risk information, so that the filtered query results are output as the execution result. Specifically, the Reduce Agent filters out data that falls within the aforementioned permission list and risk list from the query results, retains the query results that meet the permission access requirements and do not involve risky content, integrates the retained query data into the execution result, and outputs it to the MasterAgent.

[0044] It should be noted that if various permission and risk rules are pre-inputted to the agent before the task is executed, it may lead to misunderstandings or execution errors in the large model, resulting in the inability to obtain complete query data. This application places the auditing operation for the query results after the subtask query is completed. That is, the Map Agent is allowed to call the complete search content during the query to ensure the integrity of the query data. At the same time, the Reduce Agent is used to filter the query results to avoid the leakage of privacy information or unauthorized access information in the final output execution results.

[0045] This application embodiment obtains the permission information and risk information corresponding to the current task instruction through a fourth type of intelligent agent, so as to ensure that the query results comply with the data query permission regulations of different users, avoid the direct output of some user privacy or sensitive data, improve the execution security of query task results, and effectively avoid the intelligent agent's large model understanding from causing deviations or execution errors, and prevent the incomplete data from being obtained. In this way, the intelligent agent can use complete search content when querying, while preventing it from leaking privacy or sensitive information in the output.

[0046] When the task type is an operation task, step S3 above may include the following steps: Step S321: Use the first type of intelligent agent to parse the current task instruction to determine the second parsing result corresponding to the current task instruction. The second parsing result includes the first current semantic vector and the first current graph vector corresponding to the current task instruction.

[0047] Specifically, embodiments of this application utilize two modalities—a first current semantic vector and a first current graph vector—to comprehensively represent the operational requirements of the current task instruction. Specifically, a first type of agent extracts textual information from the current task instruction, such as user requests, task objectives, and problem summaries, and uses a pre-trained sentence embedding model such as Sentence-BERT to convert this textual information into a first current semantic vector.

[0048] The first type of intelligent agent also performs structured parsing of the current task instruction based on a preset knowledge base, parsing out structured information such as the resource type involved in the current task instruction, the identifier of the toolchain to be called, and the problem label to which the task belongs. Based on this structured information, it constructs a task semantic evolution subgraph (TSEG Subgraph), and then uses graph embedding algorithms such as graph neural network (GNN) or Node2Vec to encode the task semantic evolution subgraph into a first current graph vector.

[0049] Step S323: Based on the first current semantic vector and the first current graph vector, retrieve historical experience cases that match the current task instruction from the preset case memory.

[0050] Specifically, the first type of intelligent agent retrieves the k most relevant historical experience cases to the current task instruction from the preset case memory bank based on the dual-mode retrieval mechanism of the first current semantic vector and the first current graph vector. Each historical experience case corresponds to a first historical semantic vector and an experience graph vector.

[0051] In some embodiments of this application, the cosine similarity between the first current semantic vector and the historical semantic vectors of each historical experience case is first calculated to obtain the text similarity. Simultaneously, the cosine similarity between the first current graph vector and the experience graph vectors of each historical experience case is calculated to obtain the graph structure similarity. Then, combining the calculation results of the two text similarities and the graph structure similarity, the comprehensive similarity between the current task instruction and each historical experience case is determined according to a preset weight. Specifically, the comprehensive similarity is determined according to the following formula (1): In the formula, Indicates the current task instruction. Indicates historical experience cases as candidates. Indicates text similarity weights. express and Text similarity between them express and The similarity of the graph structures between them express and The overall similarity between them.

[0052] The cosine similarity mentioned above is calculated according to the following formula (2): In the formula, The result of the cosine similarity calculation ranges from [−1, 1]. The closer the result is to 1, the higher the similarity. Compared to Euclidean distance, the cosine similarity used in this application focuses more on the consistency of vector directions, which can effectively capture the semantic association between text and graph. and These are the semantic vectors or graph vectors whose similarity is to be compared, respectively. and These are the norms of the semantic vector and the graph vector, respectively. It can be understood that the text similarity is calculated by substituting the first current semantic vector and the historical semantic vector into the above formula (2), and the graph structure similarity is calculated by substituting the first current graph vector and the empirical graph vector into the above formula (2).

[0053] Then, historical experience cases whose comprehensive similarity meets the preset matching threshold are used as historical experience cases that match the current task instruction, and the retrieval operation in the preset case memory is completed.

[0054] Step S325: Generate multiple current subtasks corresponding to the current task instruction based on historical experience cases and a preset knowledge base.

[0055] Specifically, on the one hand, before formulating a computing network management strategy, the first type of intelligent agent needs to query a pre-set knowledge base to obtain relevant professional background knowledge. For example, for operation and maintenance, root cause analysis is required; for application deployment, application deployment preference modeling data needs to be obtained, such as whether the application to be deployed is compute-intensive or I / O-intensive, and relevant test benchmark data needs to be obtained. By obtaining relevant knowledge of operational requirements through the pre-set knowledge base, the operational requirements are broken down into multiple current sub-tasks to form a computing network management strategy. On the other hand, the first type of intelligent agent analyzes historical experience cases to determine whether the computing network management strategy needs optimization and adjustment. For example, if there are historical experience cases with high failure rates, potential risk points are marked, such as "GPU driver version incompatibility." If there are historical experience cases with high costs, optimization space is marked, such as "network transmission time is too long." Then, based on the marked potential risk points and optimization space, the first type of intelligent agent adjusts the already broken-down current sub-tasks. The computing network management strategy composed of the adjusted current sub-tasks includes explicit avoidance steps and priority paths, such as "checking the target node driver version" and "prioritizing containerized deployment solutions with high historical success rates."

[0056] In this way, by introducing historical experience cases into the generation process of computing network management strategies, the first type of intelligent agent can refer to past experience in executing tasks when decomposing current task instructions. This is conducive to improving the first-time execution success rate of computing network management strategies, making the generation of computing network management strategies more forward-looking, and thus improving the execution efficiency of computing network management tasks.

[0057] When the task type is an operation-type task, step S5 above may include the following steps: Step S511: Use the second type of intelligent agent to parse the current subtask to determine the third parsing result corresponding to the current subtask. The third parsing result includes the second current semantic vector and the second current graph vector corresponding to the current subtask. Specifically, similar to step S321 above, the second agent extracts the text information and structured information of the assigned current subtask, thereby generating the corresponding second current semantic vector and second current graph vector. For details, please refer to the above description, which will not be repeated here.

[0058] Step S513: Use the second type of intelligent agent to perform context analysis on the current subtask in order to determine the case retrieval conditions based on the analysis results.

[0059] Specifically, the second type of intelligent agent analyzes contextual information such as the execution stage, resource constraints, and potential risks of the current subtask. For example, if it identifies that there may be network configuration risks in the current environment, it can set corresponding search conditions so that it can search the preset case memory for historical experience cases related to the potential risks, especially historical failure cases related to network configuration risks.

[0060] Step S515: Based on the second current semantic vector, the second current graph vector, and the case retrieval conditions, search the preset case memory for historical subtasks that match the current subtask, so as to adjust the execution process of the current subtask according to the historical subtasks.

[0061] Specifically, in the ReAct loop, the second type of agent actively queries the case memory based on the aforementioned retrieval conditions, learning from past experiences to achieve dynamic and intelligent subtask execution, acquiring successful tool call sequences and lessons learned from failures. If high-scoring successful cases exist, their validated tool call sequences are reused first to avoid redundant exploration. If relevant failed cases exist, their erroneous operations are actively avoided, such as skipping commands known to trigger quota shortages.

[0062] It should be noted that the method of retrieving historical subtasks from the preset case memory based on the second current semantic vector, the second current graph vector, and the case retrieval conditions is the same as step S323 above. For details, please refer to the aforementioned explanation, which will not be repeated here.

[0063] In some embodiments of this application, step S7 may include the following steps: Step S711: During the execution of any one of the multiple current subtasks, the third type of intelligent agent receives the execution status of other subtasks in the multiple current subtasks reported by the second type of intelligent agent. Step S713: If the third type of intelligent agent detects that any of the current subtasks has an execution conflict or execution duplication with other subtasks, then the third type of intelligent agent is used to adjust the execution process of the multiple current subtasks corresponding to the second type of intelligent agent.

[0064] Step S7 may also include the following steps: Step S721: When the current subtask has finished executing, the third type of intelligent agent receives the execution result of the current subtask reported by the second type of intelligent agent in real time. Step S723: If the execution result of the current subtask is execution failure, the third type of intelligent agent generates an exception report and reports it to the first type of intelligent agent, so that the first type of intelligent agent can adjust the computing network management strategy according to the exception report and re-execute it. Step S725: If the execution result of the current subtask is successful, the third type of intelligent agent outputs the execution result of the current subtask and triggers the second type of intelligent agent to execute the next current subtask.

[0065] Specifically, for operational tasks, this application embodiment constructs task groups based on the Plan-and-Execute pattern. Compared with related technologies, roles are added and their responsibilities are adjusted within the task groups to improve the coordination of operational task execution. Specifically, the first type of intelligent agent is the operation leader, the second type is the Executor Agent, and the third type is the Coordinator Agent. The execution flow of operational tasks is described in detail below: Step 1: The user inputs the current task instruction for an operation-type task; Step 2: The Operation leader parses the current task instruction received from the user, and breaks down the current task instruction into a series of independent current subtasks based on the first parsing result; (3) Synchronize the current subtask to the Coordinator Agent; Step 3: The Coordinator Agent assigns the current subtask to the appropriate Executor Agent; Step 4: The Executor Agent independently executes the assigned current subtask. It adopts the ReAct mode, that is, after the agent infers, it calls the tool. Based on the return result of the tool, the agent infers again and decides the next action until the task is completed or an obstacle is encountered. Step 5: The Executor Agent updates the execution progress of the current subtask. It is understood that in this embodiment, the Executor Agent does not need to wait until all tasks are completed before reporting. Instead, it can proactively report updates to the Coordinator Agent at each small step completion node, so that the global task status can be continuously updated at a finer granularity. Step Six: Coordinator Agent checks progress: During the execution of subtasks by the Executor Agent, the Coordinator Agent checks whether there are duplicates or conflicts in the current subtasks that the Executor Agent is waiting to execute or is currently executing. If so, the Coordinator Agent will coordinate with each Executor Agent to cancel the Executor Agent from executing the duplicate or conflicting current subtasks. When an Executor Agent completes its current subtask, it needs to report the execution result to the Coordinator Agent, regardless of whether the execution result is successful or not. This includes the subtask completion status, any problems encountered, and error messages. The Coordinator Agent checks the execution results reported by the Executor Agent. If the execution is successful, the Coordinator Agent updates the task execution status and then returns to step four to continue scheduling the next current subtask. If the execution fails, the Coordinator Agent summarizes and processes the execution status and failure error information reported by the Executor Agent to generate an exception report and reports the exception report to the Operation Leader, thereby returning to step two to regenerate the computing network management policy.

[0066] Furthermore, some embodiments of this application introduce the Agent2Agent (A2A) protocol to standardize interactions between intelligent agents. Each intelligent agent declares its capabilities and interfaces by publishing standardized agent cards, supporting automatic discovery and dynamic collaboration. During task execution, intelligent agents communicate via structured messages through the A2A protocol, supporting synchronous requests, asynchronous streaming responses, and event notifications to improve collaboration efficiency. Referring to the A2A security design, communication between intelligent agents employs OAuth2.0 authentication and mTLS encryption to ensure identity trust and data security. Access scopes are allocated according to the principle of least privilege to prevent unauthorized operations, and complete audit logs are recorded to ensure the controllability and traceability of the computing network scheduling process.

[0067] Furthermore, the preset case memory is constructed based on historical task instructions and multiple execution processes of those instructions. In a computing network scheduling and management platform based on a multi-agent architecture, multiple agents collaborate to complete complex resource scheduling tasks. Each agent generates a series of operation trajectories during task execution, including resource querying, task orchestration, fault handling, and strategy adjustment. In related technologies, the raw data of these operation trajectories are typically recorded only for auditing or monitoring purposes, lacking a structured accumulation and reuse mechanism for experience. Therefore, in this embodiment, the preset case memory is constructed using the architecture of the Hierarchical Case Memory Repository (HCMR) shown in Table 1.

[0068] Table 1 Specifically, the embodiments of this application construct a preset case memory library in the following manner: Step 1: Generate raw execution logs based on the multiple execution processes of historical task instructions.

[0069] Specifically, the L1 layer described above is constructed based on the original execution logs. This layer is designed to completely and immutably record every step of each agent's operation during task execution, and it has the following characteristics: (1) Write-only: Once a record is written, it cannot be modified; (2) High-fidelity context: includes context information such as tool call details, parameters, return values, execution time and environment snapshot.

[0070] (3) Data format: Structured JSON logs are used for easy subsequent parsing.

[0071] Step 2: Generate a first historical semantic vector based on the textual features of the historical task instructions. Determine multiple historical subtasks based on each execution process of the historical task instructions, and generate a second historical semantic vector based on the textual features of the historical subtasks. Generate entity nodes based on the historical subtasks, and generate edges based on the dynamic relationships between the entity nodes, so as to construct multiple historical task evolution graphs corresponding to the historical task instructions based on the entity nodes and edges.

[0072] Specifically, textual features of historical task instructions and subtasks are extracted using sentence embedding models such as Sentence-BERT and encoded into corresponding historical semantic vectors. Simultaneously, the aforementioned L2 layer is constructed based on the historical task evolution graph. This layer primarily aims to build the Task Semantic Evolution Graph (TSEG) to represent the dynamic relationships between task operations. The main components of the historical task evolution graph include entity nodes and edges. A triple dynamic edge mechanism supports more refined semantic analysis. This embodiment of the application supports more refined semantic analysis through a triple dynamic edge mechanism.

[0073] The aforementioned entity nodes include tasks, agents, resources (nodes / links), tools, etc. The dynamic relationships represented by the aforementioned edges are as follows: (1) Temporal Edge: Represents the time interval in which an event occurs, supporting the modeling of periodic tasks; (2) Causal Edge: Represents the causal relationship between action and result inferred from a lightweight model; (3) Dependency Edge: Represents the hard / soft dependency relationship between tools or resources.

[0074] The process of constructing the historical mission evolution map is as follows: (1) Entity extraction steps: Extract entities such as tasks, tools, resources, and states from the L1 level; (2) Relationship inference steps: Based on the timestamp information corresponding to each entity during the task execution process, the temporal relationship between entities is automatically established; combined with the preset causal judgment rules and lightweight models such as logistic regression, the causal relationship between entities is analyzed and judged; based on the tool call chain and resource consumption formed during the task execution process, the dependency relationship between entities is generated. (3) Graph update steps: The asynchronous processing method is adopted to add the newly added entity nodes and edges during the task execution to the constructed graph structure, while the historical evolution path of the graph is fully preserved to ensure that the graph can continuously reflect the changes in the entire process of task execution.

[0075] Step 3: Scan the historical task evolution map based on a preset period to obtain the empirical map vector based on the scanning results. The empirical map vector represents the map sub-path that appears most frequently in each historical task evolution map.

[0076] Specifically, the L3 layer is constructed based on the experience graph vector. This layer is mainly composed of the Experience Extraction Agent (EEA) which automatically identifies high-frequency and high-value paths from the TSEG and generates structured experience units (EUs).

[0077] EEA is a dedicated intelligent agent based on LLM, and its workflow is as follows: (1) Pattern Detection Steps: Regularly scan TSEG and use graph matching algorithms to discover repetitive structures in the graph, thereby identifying graph sub-paths that appear frequently and have high success rates in the evolution graphs of each historical task. (2) Semantic Summarization Step: Input the identified graph sub-paths and their tool call sequences, state transitions, environmental conditions and other contextual information into the LLM model. Examples of prompt words are as follows: "You are an expert in the field of network scheduling. Please summarize the following operation trajectory into a structured experience unit."

[0078] Require: Extracting problem patterns and objectives List the complete tool execution steps (including parameters and expected results). Annotate applicable scenarios and constraints Output in JSON format [Input: Multiple similar paths extracted from TSEG and their context] (3) Structured output steps: LLM outputs standardized empirical units, which contain the following key fields: a.solution_path: Maintains a complete tool call chain, including parameters, expected output, and success conditions, to ensure executability; b. experience: Record typical errors encountered in the actual implementation of the solution, root cause analysis, coping strategies and lessons learned, to enhance the practical guidance value of the case; c.related_cases: Lists other cases that are similar to this case in terms of problem pattern, solution or technology stack, supporting cross-case knowledge association and recommendation; d.source_traces: Explicitly marks the original trajectory IDs on which this case was generated, ensuring traceability; (4) Integrity verification steps: The system automatically checks whether all tools are available in the current environment, whether the parameters are complete, and whether they contain success criteria; if the verification fails, it returns modification suggestions and regenerates the experience units. If the verification passes, the verified experience units are converted into structured experience map vectors.

[0079] It is understood that the embodiments of this application represent the execution process of historical task instructions and their corresponding historical subtasks through experience graph vectors, thereby providing experience for task decomposition of the first type of intelligent agent and task execution of the second type of intelligent agent.

[0080] Step 4: Generate historical experience cases based on the first historical semantic vector, the second historical semantic vector, the experience graph vector, historical subtasks, and historical task instructions, and build a preset case memory bank based on the historical experience cases.

[0081] Therefore, this application constructs a raw trajectory memory layer by generating raw execution logs to facilitate the tracing of complete data on the execution process of historical task instructions. A semantic association memory layer is constructed by generating historical semantic vectors and historical task evolution graphs to deeply represent the dynamic relationships such as the temporal sequence, causality, and tool or resource dependencies of each operation during execution. A case generation layer is constructed by extracting experience graph vectors from the historical task evolution graph, thereby forming historical experience cases in the form of structured data. In this way, by constructing a traceable and structured hierarchical case library, this application provides an effective experience foundation for the generation of network management strategies in subsequent operational tasks, thereby further improving the execution efficiency of network management tasks.

[0082] It should be noted that the aforementioned preset case memory is used, on the one hand, to guide the generation of the Operation Leader's computing network management strategy, and on the other hand, to guide the execution of the current subtask of the Executor Agent.

[0083] In some embodiments of this application, as the computing network scheduling platform operates, memory records in the preset case memory bank are continuously generated. If there is a lack of an effective memory update mechanism, invalid or expired memories will occupy a large amount of storage resources, affecting system performance. Therefore, the preset case memory bank is updated in the following way: Step 1: If the current subtask has finished executing, take the execution process of the current subtask as the current case, and determine the value score of the current case based on the execution result of the current subtask.

[0084] Furthermore, the value score of the current case is determined based on the execution result of the current subtask, including: An execution quality index is determined based on the execution result of the current subtask, and the execution quality index represents the execution completeness and success of the current subtask; The execution cost is determined based on the execution time and resource consumption of the current subtask; The experience reward for the current subtask is determined based on the problem-solving situation during the execution of the current subtask. The execution quality indicators, execution costs, and experience rewards are weighted and calculated based on preset weights to determine the value score according to the calculation results.

[0085] Optionally, the value score of the current case can be calculated according to the following formula (3): In the formula, This represents the value score of the current case c. and To achieve the quality targets, among which, The integrity of execution is characterized by comparing each step in the execution process with an expert guide to the task execution flow in a pre-defined knowledge base. The more each step in the execution process conforms to the requirements of the expert guide, the better. The higher the value, the better. This indicates the success rate of execution; the more successful steps during execution, the better. The higher the value, the better. The execution cost is evaluated based on the tokens consumed in task execution and the time required to complete the task. As an experience reward, among which, This is the set of problems encountered during the execution of the current case c. This is the difficulty level of the problem; the more difficult the problem, the higher the set value. This indicates whether the problem has been resolved; a value of 1 indicates a resolved problem, and a value of 0 indicates an unresolved problem. Rate the user's feedback on the current case. If the user deems the current case (c) worthy of attention, then rate it. Bonus points will be awarded accordingly to meet users' customization needs for the case studies. , , and The weights for execution quality indicators, execution costs, experience rewards, and feedback scores are listed in order. and In order and The corresponding sub-weights.

[0086] This application embodiment scores the value of a current case by comprehensively evaluating dimensions such as the completeness of the plan, success rate, and execution cost. This value score can simultaneously reflect the task execution effect and execution cost, thereby providing an accurate and reasonable basis for subsequent updates to the case library based on value scores.

[0087] Step 2: Based on the first current semantic vector and the first current graph vector, retrieve relevant experience cases from the preset case memory. The relevant experience cases are the top k historical experience cases with the highest similarity to the current case.

[0088] Specifically, the comprehensive similarity between the current case and each historical experience case in the preset case memory bank is determined based on the above formula (1). The calculation formula is the same as that of formula (1) above. Please refer to the above explanation for details. It should be noted that in this step, the formula (1) This represents the current case. This results in a relevant case set, SimSet, consisting of the k most similar historical cases.

[0089] Step 3: Determine the adaptive update threshold corresponding to the current case. The adaptive update F threshold is determined based on the importance level of the operation task, the current capacity of the preset case memory, and the value score distribution of relevant experience cases.

[0090] Specifically, the above adaptive update threshold is determined based on the following formula (4): In the formula, This indicates an adaptive threshold update. This indicates the current capacity of the preset case memory, i.e., the current number of cases. This indicates the preset high load threshold of the case library. This indicates a subclass belonging to the operation task category, where critical represents critical tasks, security represents security tasks, routine represents routine operation tasks, and optimization represents optimization tasks. The score cutoff for the top 10% of categories k. The score cutoff for the top 50% of categories k. The score line representing the bottom 20% of the scores in category k.

[0091] Therefore, formula (4) above can be used to achieve differentiated configuration of case selection strategies for different types of tasks. For critical tasks, the adaptive update threshold is set at a higher level to ensure that the case entry standard remains at a high level. For routine tasks, the case entry standard can be appropriately lowered to enrich the case diversity corresponding to routine tasks while ensuring the effectiveness of the cases.

[0092] Step 4: If the value score of the current case is greater than the adaptive update threshold, update the preset case memory based on the current case.

[0093] Specifically, if If so, the current case c will be inserted into the preset case memory.

[0094] This application employs a multi-agent aggregated case adaptive update (MACAU) algorithm to retrieve relevant experience cases. Combined with adaptive update thresholds determined from multiple dimensions, it ensures that high-value current cases are inserted into the preset case memory to update the preset case memory, preventing invalid or expired memories from occupying a large amount of storage resources and affecting system performance.

[0095] Furthermore, after updating the preset case memory, this embodiment of the application also updates the score of the relevant case set SimSet, including: The score change is determined based on the similarity between relevant experience cases and the current case, the preset update intensity, and the average value score of relevant experience cases, so as to update the scores of relevant experience cases and the current case according to the score change.

[0096] Specifically, the change in score is calculated based on the following formula (5): In the formula, This indicates the change in rating. Indicates the strength of the rating update. For indicator functions, The baseline representing category k is determined by the mean value scores of relevant experience cases in the SimSet set of related cases. The average value score for each relevant experience case and the current case.

[0097] Then, the corresponding scores are updated for each relevant experience case and the current case in the relevant case set SimSet, as shown in the following formula (6): In the formula, This indicates the value rating before the update. This indicates the updated value rating.

[0098] Furthermore, this application embodiment also draws analogy with human memory, updating the scores of each historical experience case in the preset case memory bank based on time decay, as shown in the following formula (7): In the formula, The value score of historical experience case c after time decay is updated. This represents the value score of historical case c before time decay update. Indicates the attenuation rate. Indicates a unit of time. This indicates the creation time of historical experience case c. This indicates the number of times historical experience case c has been reused. This represents the enhancement coefficient.

[0099] It should be noted that the natural forgetting part of formula (7) is based on the Ebbinghaus forgetting curve of human memory, while the reinforcement part considers the number of times historical experience cases are used, so as to realize a dynamic update mechanism of high-frequency use reinforcement and low-frequency use decay of cases.

[0100] In some embodiments of this application, if the capacity of the preset case memory is... Exceeding the upper limit, i.e., exceeding the overflow threshold. If a historical experience case c satisfies both formula (8) and formula (9) below, then the historical experience case c is deleted from the preset case memory. In the formula, , This represents the average value score of all cases in the current preset case memory. This represents the standard deviation of the value scores for all cases in the current preset case memory. This indicates the minimum number of days a case must be retained. Indicates the current time. Indicates the creation time of the case.

[0101] This application embodiment achieves dynamic optimization and continuous evolution of the preset case memory by updating the scores of similar cases, ensuring the high quality and timeliness of the preset case memory.

[0102] Accordingly, please refer to Figure 2 This application provides a computer network intelligent management device, which includes: The task group creation unit 100 is used to determine the task type in response to the current task instruction and to determine the task group that matches the task type. The task group includes a first type of intelligent agent, a second type of intelligent agent and a third type of intelligent agent. For details, please refer to the description of step S1 above. The strategy generation unit 200 is used to generate a computing network management strategy using the first type of intelligent agent according to the current task instructions. The computing network management strategy includes multiple current sub-tasks, as detailed in the description of step S3 above. The execution unit 300 is used to execute each current subtask using the second type of intelligent agent and output the execution result of the current subtask. For details, please refer to the description of step S5 above. The coordination unit 400 is used to send the execution status of the current subtask to the third type of intelligent agent in real time when the task type is an operation task. This allows the third type of intelligent agent to dynamically adjust the execution of the current subtask based on the execution status during the execution process, or to trigger the first type of intelligent agent to dynamically adjust the computing network management strategy based on the execution status after the current subtask is completed. For details, please refer to the description of step S7 above.

[0103] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0104] In this embodiment, the intelligent management device of the computing network is presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0105] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application, such as... Figure 3As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 3 Take a processor 10 as an example.

[0106] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0107] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0108] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0109] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0110] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0111] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the methods shown in the above embodiments are implemented.

[0112] This application provides a computer program product including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method of any embodiment of this application.

[0113] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

[0114] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0115] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this 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, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0117] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0118] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0119] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0120] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0121] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0122] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for intelligent management of computing networks, characterized in that, The method includes: In response to the current task instruction, a task type is determined, and a task group matching the task type is determined, the task group including a first type of intelligent agent, a second type of intelligent agent, and a third type of intelligent agent; The first type of intelligent agent generates a computing network management strategy based on the current task instruction, and the computing network management strategy includes multiple current sub-tasks; The second type of intelligent agent is used to execute each of the current sub-tasks and output the execution results of the current sub-tasks; When the task type is an operation task, the second type of intelligent agent sends the execution status of the current sub-task to the third type of intelligent agent, so that the third type of intelligent agent can adjust the current sub-task according to the execution status during the execution of the current sub-task, or trigger the first type of intelligent agent to dynamically adjust the computing network management strategy according to the execution status after the current sub-task is completed.

2. The intelligent management method for computing networks according to claim 1, characterized in that, When the task type is a query task, the step of generating a network management strategy using the first type of intelligent agent based on the current task instruction includes: The first type of intelligent agent is used to parse the current task instruction to determine the first parsing result corresponding to the current task instruction; Based on the first parsing result, multiple current subtasks are determined in parallel.

3. The method according to claim 2, characterized in that, The method further includes determining a fourth type of intelligent agent that matches the query-type task; the method further includes: After all the current subtasks have been executed, the fourth type of intelligent agent is used to obtain the permission information and risk information corresponding to the current task instruction; The third type of intelligent agent is used to obtain the query results of each current subtask, and the query results are filtered according to the permission information and the risk information, so as to output the filtered query results as the execution result.

4. The intelligent management method for computing networks according to claim 1, characterized in that, When the task type is an operational task, the step of generating a network management strategy using the first type of intelligent agent based on the current task instruction includes: The first type of intelligent agent is used to parse the current task instruction to determine the second parsing result corresponding to the current task instruction. The second parsing result includes the first current semantic vector and the first current graph vector corresponding to the current task instruction. Based on the first current semantic vector and the first current graph vector, historical experience cases matching the current task instruction are retrieved from the preset case memory. Based on the historical experience cases and the preset knowledge base, generate multiple current subtasks corresponding to the current task instruction; The preset case memory is constructed based on historical task instructions and the multiple execution processes of those historical task instructions.

5. The intelligent management method for computing networks according to claim 4, characterized in that, The historical experience cases stored in the preset case memory bank include multiple historical sub-tasks; The execution of each of the current sub-tasks using the second type of intelligent agent includes: The second type of intelligent agent is used to parse the current subtask to determine the third parsing result corresponding to the current subtask. The third parsing result includes the second current semantic vector and the second current graph vector corresponding to the current subtask. The second type of intelligent agent is used to perform context analysis on the current subtask in order to determine the case retrieval conditions based on the analysis results; Based on the second current semantic vector, the second current graph vector, and the case retrieval conditions, the historical subtasks that match the current subtask are retrieved from the preset case memory, so as to adjust the execution process of the current subtask according to the historical subtasks.

6. The method according to claim 5, characterized in that, The method constructs the preset case memory library in the following manner: A first historical semantic vector is generated based on the textual features of the historical task instructions; Multiple historical subtasks are determined based on each execution process of the historical task instruction, and a second historical semantic vector is generated based on the textual features of the historical subtasks. Entity nodes are generated based on the historical sub-tasks, and edges are generated based on the dynamic relationships between the entity nodes, so as to construct multiple historical task evolution graphs corresponding to the historical task instructions based on the entity nodes and the edges; The historical task evolution map is scanned based on a preset period to obtain an empirical map vector based on the scanning results. The empirical map vector represents the map sub-path that appears most frequently in each of the historical task evolution maps. Historical experience cases are generated based on the first historical semantic vector, the second historical semantic vector, the experience graph vector, the historical subtask, and the historical task instruction, so as to construct the preset case memory bank based on the historical experience cases.

7. The method according to claim 1, characterized in that, The step of adjusting the execution of the current subtask based on the execution status includes: During the execution of any of the plurality of current subtasks, the third type of intelligent agent receives the execution status of other subtasks among the plurality of current subtasks reported by the second type of intelligent agent; If the third type of intelligent agent detects that any of the current subtasks has an execution conflict or duplicate execution with the other subtasks, then the third type of intelligent agent is used to adjust the execution process of the multiple current subtasks corresponding to the second type of intelligent agent.

8. The method according to claim 1, characterized in that, The step of triggering the first type of intelligent agent to dynamically adjust the computing network management strategy based on the execution status includes: When the current subtask has finished executing, the third type of intelligent agent receives the execution result of the current subtask reported by the second type of intelligent agent in real time. If the execution result of the current subtask is an execution failure, the third type of intelligent agent generates an exception report and reports it to the first type of intelligent agent, so that the first type of intelligent agent can adjust the computing network management strategy according to the exception report and re-execute it; If the execution result of the current subtask is successful, the third type of intelligent agent outputs the execution result of the current subtask and triggers the second type of intelligent agent to execute the next current subtask.

9. The method according to claim 4, characterized in that, The method also includes updating the preset case memory in the following ways: When the current subtask has finished executing, the execution process of the current subtask is taken as the current case, and the value score of the current case is determined based on the execution result of the current subtask. Based on the first current semantic vector and the first current graph vector, relevant experience cases are retrieved from the preset case memory. The relevant experience cases are the top k historical experience cases with the highest similarity to the current case. Determine the adaptive update threshold corresponding to the current case, wherein the adaptive update threshold is determined based on the importance level of the operation task, the current capacity of the preset case memory, and the value score distribution of the relevant experience cases; If the value score of the current case is greater than the adaptive update threshold, the preset case memory is updated according to the current case.

10. The method according to claim 9, characterized in that, Determining the value score of the current case based on the execution result of the current subtask includes: An execution quality index is determined based on the execution result of the current subtask, and the execution quality index represents the execution completeness and success of the current subtask; The execution cost is determined based on the execution time and resource consumption of the current subtask; The experience reward for the current subtask is determined based on the problem-solving situation during the execution of the current subtask. The execution quality indicators, execution costs, and experience rewards are weighted and calculated based on preset weights to determine the value score according to the calculation results.

11. The method according to claim 9, characterized in that, After updating the preset case memory, the method further includes: The score change is determined based on the similarity between the relevant experience cases and the current case, the preset update intensity, and the average value score of the relevant experience cases, so as to update the scores of the relevant experience cases and the current case according to the score change.

12. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected and the memory stores computer instructions. The processor executes the computer instructions to perform the intelligent management method of any one of claims 1 to 11.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the intelligent management method of the computer network as described in any one of claims 1 to 11.

Images