Cyclic feedback scheduling cooperation method based on agent and cloud computing power service

Abstract

The invention provides a cyclic feedback scheduling collaboration method based on an intelligent agent and cloud computing power service. The method comprises the following steps: memory recall: the intelligent agent receives a task demand, accesses a memory library to obtain historical experience entries, performs memory matching and memory fusion, generates memory recall information, and outputs the memory recall information to the cloud computing power service; cloud task planning: a cloud big language model understands task demand intentions based on memory recall information, decomposes tasks and generates executable descriptions, and an agent generates a preliminary task plan and calls a proper tool; performing action observation: performing operation by the intelligent agent, acquiring feedback data in real time, comparing an actual result with an expected target adjustment strategy, selecting a more suitable tool, and performing repeated execution until a task is completed; summarizing and reflecting: generating task execution tracks, integrating into experience entries, and storing into a memory bank. According to the method, historical memory, dynamic feedback and cloud computing power are combined, and the autonomous decision-making ability, execution efficiency and accuracy of the intelligent agent in a complex task are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of artificial intelligence, large language models and cloud computing, and specifically relates to a cyclic feedback scheduling cooperation method based on an agent and cloud computing power service. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, agents are increasingly widely used in complex task processing, especially in task scheduling, resource coordination and dynamic decision-making. In recent years, large language model (LLM) technology has developed rapidly, and with its powerful semantic understanding ability and natural language interaction interface, it can support task planning and execution of agents to some extent, enhancing the task processing ability of agents. As a direct landing of large language model technology, the application potential of agents in complex task environments gradually emerges. However, traditional agent algorithms usually rely on a large amount of training data or frequent trial-and-error optimization strategies, which not only has high cost and long cycle, but also has certain limitations in dynamic environments. In the application scenario of LLM-enabled agents, although LLM has extensive internal world knowledge and the ability to interact with users through a natural language interface, agents may experience "model hallucination" when planning and making decisions in complex tasks due to insufficient understanding or lack of knowledge. This phenomenon is particularly evident when there is a lack of specific task training or when faced with complex and variable environments, resulting in low efficiency and frequent errors in task execution. In addition, existing agents lack effective use of historical experience during task execution, leading to frequent repetitive errors and inefficient attempts.

[0003] To address the above challenges, researchers have proposed various improvement methods, such as implementing self-refinement through iterative frameworks, optimizing prompt writing using gradient descent and beam search techniques, and learning from environmental feedback based on policy gradients. However, these methods are often limited to single-step generation tasks or specific domains, making it difficult to handle complex multi-step tasks across domains. Existing technologies, such as Chinese patent publication CN118132227A, disclose a method for agent scheduling using a large language model. This method decomposes task execution instructions into multiple subtasks and determines the execution order of subtasks and the required agents by a large language model, thereby improving the accuracy and efficiency of task execution. However, this method still has limitations in handling complex multi-step tasks across domains, and relies on fine-tuning of the large language model, which is costly and limits its widespread application. SUMMARY

[0004] In order to solve the problems of the agent in the prior art in the complex task, such as relying on trial and error optimization, insufficient dynamic environment adaptability, and decision deviation of large language model caused by "model hallucination", a circulating feedback scheduling cooperation method based on agent and cloud computing power service is provided, and through dynamically fusing local decision of agent, cloud computing power resource and historical experience feedback, the dual improvement of task execution efficiency and accuracy is realized.

[0005] The application is implemented as follows: a circulating feedback scheduling cooperation method based on agent and cloud computing power service, comprising the following steps:

[0006] Step S1, memory recall: after the agent receives the task demand, the memory bank is accessed to obtain the historical experience entries, and memory matching and memory fusion are performed locally in combination with the task demand to generate memory recall information, and the memory recall information is output to the cloud computing power service for analysis, and the cloud computing power service includes a cloud large language model;

[0007] Step S2, cloud task planning: the cloud large language model performs intention understanding, task decomposition on the task demand based on the memory recall information generated in step S1, and generates an executable description, and the content of the executable description includes execution logic, dependency relationship and execution target of the task; the agent generates a preliminary task plan according to the executable description and calls tools meeting the demand from the tool set;

[0008] Step S3, action observation: the agent performs operations and observes the results, collects feedback data of the environment and the used tools in real time, compares the actual results with the expected target, adjusts the task strategy, selects more suitable tools from the tool set for subsequent operations; the agent updates and circulates the task, repeatedly executes this step until it is determined that the task is finally completed;

[0009] Step S4, summary and reflection: the agent summarizes and reflects on the whole execution process of step S3, records and generates a task execution track, and integrates it into an experience entry and stores it in the memory bank in step S1, and the experience entry includes execution record, tool selection, operation step, problem encountered and corresponding solution.

[0010] Further, the formula of memory matching in step S1 is:

[0011]

[0012] In the formula, q is an input vector of the current task; m i is a memory vector of the i-th historical experience entry in the memory bank; sim(q, m i) is cosine similarity, the most similar historical experience entry to the current task q is selected by calculating the cosine similarity, when the cosine similarity is higher than a set threshold, it means that the historical experience entry has strong relevance with the current task, which can be used for memory fusion in step S1.

[0013] Further, the set threshold value range is 60%-80%.

[0014] Further, the formula of memory fusion in step S1 is:

[0015] S i =sim(q,m i )·λ i

[0016] In the formula, λ i is a logical level relevance weight, which is introduced on the basis of the cosine similarity, representing the logical relevance degree of the i-th historical experience entry m i in the current task q; S i is the comprehensive relevance score.

[0017] Further, the logical level relevance weight λ i is adaptively adjusted by the cloud large language model according to the task requirements, and the value range is [0, 1].

[0018] Further, the intent understanding and task decomposition of the task requirement in step S2 are specifically: assuming that the current task is T, the intent understanding result vector of the task is I(T), and the memory recall information generated in step S1 is C, then the formula of intent understanding and task decomposition is:

[0019] I(T)=LLM intent (T)

[0020] ST i =LLM decompose (I(T),C i )

[0021] In the formula, LLM intent (·) is the intent analysis ability function of the cloud large language model; C i is the context information used for decomposing the i-th subtask; LLM decompose (·) is the task decomposition function of the cloud large language model; ST i is the i-th subtask.

[0022] Further, the formula for generating an executable description in step S2 is:

[0023] Exec_Desc′(T)=LLM generate_description{ST1,ST2,...,ST N})

[0024] wherein {ST1,ST2,...,ST N} is a subtask sequence of the current task T; LLM generate_descriotion (·) is a task induction capability function of the cloud large language model; Exec_Desc'(T) is an executable description of the current task T.

[0025] Further, the determination that the task is finally completed in step S3 is that all subtasks are completed.

[0026] Further, the formula for generating the task execution trajectory in step S4 is:

[0027] Exec_Log(T)={[ST1,Tool1,O1],[ST2,Tool2,O2],...,[ST N ,Tool N ,O N ]}

[0028] wherein ST1,ST2,...,ST N are subtasks of the current task T; Tool1,Tool c ,...,Tool N are tools used to complete the subtasks; O1,O2,...,O N are observed execution results of the subtasks; and Exec_Log(T) is an execution trajectory of the current task T.

[0029] The beneficial effects of the present application are: (1) combined with historical memory, dynamic feedback and optimized scheduling of cloud computing resources, through a cyclic feedback scheduling coordination mechanism, the autonomous decision-making ability, task execution efficiency and accuracy of the agent in a complex task environment are significantly improved. (2) Make full use of cloud computing power services for efficient computing and task analysis, reduce the dependence of the agent on large computing power and frequent trial-and-error optimization strategies, and reduce the cost and time input. (3) Through a detailed summary and reflection mechanism, the task processing ability of the agent is continuously optimized, its adaptability in a dynamic environment is enhanced, the "model illusion" phenomenon is avoided, and the robustness and practicality of the algorithm are improved.

[0030] The present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a flowchart of the present application;

[0032] Figure 2 is a whole structure diagram of embodiment one of the present application. DETAILED DESCRIPTION

[0033] Example 1:

[0034] This embodiment provides a loop feedback scheduling collaboration method based on intelligent agents and cloud computing services, such as Figure 1 As shown, the following steps are included:

[0035] Step S1, memory recall:

[0036] After receiving a task, the agent accesses its memory bank to retrieve historical experience items. It then performs local memory matching and fusion based on the task requirements to generate memory recall information. This memory recall information is then exported to a cloud computing service for analysis, which includes a large cloud language model.

[0037] The formula for memory matching is:

[0038]

[0039] Let q be the input vector of the current task, m i is the memory vector of the i-th historical experience item in the memory bank, then the historical experience item that is most similar to the current task q can be selected by calculating the cosine similarity. i ) is above the set threshold, indicating that the historical experience entry is strongly relevant to the current task and can be used for memory fusion, providing rich decision support for cloud computing services. The cosine similarity threshold ranges from 60% to 80%. In this embodiment, 80% is used as the threshold.

[0040] For the input vector of the current task and each recalled memory vector, a logical hierarchical correlation weight is introduced based on the cosine similarity to form a comprehensive correlation score, that is, the memory fusion formula is used:

[0041] S i =sim(q,m i )·λ i

[0042] Where λ i is the logical level relevance weight, representing the i-th historical experience item m i The degree of logical relevance in the current task q can be adaptively adjusted by the cloud-based large language model according to the requirements of the task; S i is the comprehensive relevance score. Weight λ i The value range of is [0, 1]. In this embodiment, the initial value of the weight is set to 0.6, and it is adaptively adjusted according to the task requirements starting from the initial value of 0.6.

[0043] Step S2, cloud task planning:

[0044] The cloud large language model performs intent understanding and task decomposition on the memory recall information generated in step S1, and generates an executable description. Let the current task be T, the intent understanding result vector of the task be I(T), and the memory recall information generated in step S1 be C. The formula for intent understanding and task decomposition is:

[0045] I(T) = LLM intent (T)

[0046] ST i = LLM decomoose (I(T), C i )

[0047] In the formula, LLM intent (·) is the intent analysis capability function of the cloud large language model; C i is the context information used when decomposing the i-th subtask, such as task requirements, constraints, and historical memory; LLM decompose (·) is the task decomposition function of the cloud large language model; and ST i is the i-th subtask. After task decomposition by the cloud large language model, the current task T is decomposed into N subtasks {ST1, ST2,..., ST N}, which will serve as the basis for further planning.

[0048] After decomposing the subtasks, the cloud large language model generates an executable description, with the formula being:

[0049] Exec_Desc'(T) = LLM generate_description ({ST1, ST2,..., ST N})

[0050] In the formula, LLM generate_description (·) is the task induction capability function of the cloud large language model; and Exec_Desc'(T) is the executable description of the current task T, which is generated by the cloud large language model according to the subtask sequence {ST1, ST2,..., ST N}. The content of the executable description includes the execution logic, dependency relationship, and execution target of the task. Subsequently, the agent generates a preliminary task plan according to the executable description and calls tools that meet the requirements from the tool set.

[0051] Step S3, action observation:

[0052] The intelligent agent performs operations and observes the results, collects feedback data of the environment and the tools used in real time, compares the actual results with the expected target, adjusts the task strategy, and selects more suitable tools from the tool set for subsequent operations. The intelligent agent updates and cycles the task, repeatedly performs this step, and determines that the final task is completed when all sub-tasks are completed.

[0053] Specifically, the cloud large language model divides a whole task into multiple sub-tasks through task planning, and matches appropriate execution tools for each sub-task. Each sub-task is regarded as a single node, and the accuracy of calling tasks by the single node is evaluated. Calling tasks refer to selecting tools and executing (deciding first, then calling), and sub-tasks are matched with tools. According to the evaluation of this embodiment, the accuracy of calling tasks by the single node can reach 95% in the process of executing all sub-tasks. At the same time, the large language model dynamically adjusts the task flow according to the execution results of the sub-tasks, and continuously optimizes until the whole task is completed.

[0054] Step S4, summary and reflection:

[0055] The intelligent agent summarizes and reflects on the whole execution process of step S3, records and generates a complete task execution track. The task execution track Exec_Log(T) of the current task T is generated according to the following formula:

[0056] Exec_Log(T)={[ST1,Tool1,O1],[ST2,Tool2,O2],...,[ST N ,Tool N ,O N ]}

[0057] In the formula, ST1, ST2,..., ST N are sub-tasks of the current task T; Tool1, Tool2,..., Tool N are tools used to complete the sub-tasks; O1, O2,..., O N are observed execution results of the sub-tasks.

[0058] Finally, the intelligent agent integrates the execution records, tool selection, operation steps, problems encountered and corresponding solutions into experience entries and stores them in the memory library. The goal of this stage is to form experience memory through detailed summary, provide reference information for future execution of similar tasks, and continuously optimize the task processing ability of the intelligent agent.

[0059] Embodiment two:

[0060] This embodiment is a specific application of embodiment one. This embodiment develops an intelligent agent-based information question and answer application using the method provided in embodiment one.

[0061] (1) Memory recall stage. After receiving the user's question, the agent first retrieves relevant historical question records from the local memory library, and according to the current task requirements, it performs semantic matching and content fusion processing on the retrieved memory fragments to generate a preliminary memory recall result. Subsequently, the result is transmitted to the cloud computing platform for in-depth analysis. The large language model on the platform performs multi-dimensional reasoning, context completion, and knowledge enhancement on the memory information, and finally outputs an optimized intelligent response content.

[0062] (2) Cloud task planning stage. The cloud large language model first performs multi-dimensional intent analysis on the user's input task requirements based on the context information provided by the memory recall module, such as user historical query records, domain knowledge graph associated data, etc. This includes identifying implicit information categories, timeliness requirements, and output format preferences. For example, the information category can be real-time news, technical white paper, industry report, the timeliness can choose the latest information or historical information summary, and the output format can be abstract, information chart, bilingual translation comparison, etc.

[0063] Then, the model decomposes complex requirements into a structured sub-task chain through a task decomposition engine, and generates executable description instructions for each sub-task. For example, "information retrieval, semantic filtering, format conversion, multi-modal presentation" are four sub-tasks. The agent generates a preliminary task plan based on the executable description instructions and automatically calls the cloud tool set that meets the requirements through the dynamic tool adaptation module, realizing the user's input requirements and the agent's output of accurate information.

[0064] (3) Action observation stage. The agent realizes dynamic task optimization through the execution-observation-iteration mechanism. When executing the task operation, the agent collects real-time environmental feedback data and tool output results, and performs multi-dimensional comparative analysis based on the preset target. If it finds that the actual result deviates from the target, the agent will automatically trigger strategy adjustment: re-evaluate the tool adaptability through reinforcement learning algorithm, and dynamically select a more suitable tool combination.

[0065] For example, the agent calls the news API to grab data and the NLP model to generate an abstract, and collects real-time environmental data such as API response delay, data source update frequency, etc., as well as indicators such as tool output abstract keyword coverage, chart visualization effect, etc. Based on the preset target "information timeliness ≤ 24 hours" "abstract accuracy ≥ 95%", if the news timeliness is over-standard or the abstract misses core information, etc., automatically switch to a news source with higher real-time performance, adjust the parameter weight of the NLP model; superimposed use of public opinion analysis tools to supplement background information, call multi-language translation engine to increase foreign language information, and realize high-quality information output.

[0066] (4) Summary and reflection stage. In the above example, the agent records the task cycle data as follows:

[0067] Tool chain: News API -> NLP summarization model -> visualization engine

[0068] Operation sequence: keyword filtering -> multi-source data aggregation -> timeliness verification

[0069] Environment parameters: API response delay, data source update frequency

[0070] Based on preset rules and machine learning models, perform deep attribution analysis on execution records:

[0071] Identify tool adaptability bottlenecks: 40% of a news source's response time exceeds in sudden scenarios

[0072] Operation step redundancy point: repeated execution of 3 invalid keyword expansion

[0073] Result quality defects: 20% of the core policy text expression is omitted in the summary

[0074] Optimization strategy: switch to backup API + add caching mechanism, simplify keyword expansion rules

[0075] Implementation effect: response time is shortened by 60%, summary accuracy is improved to 95%.

[0076] All the above information is encapsulated as reusable experience entries, dynamically stored in the memory library and associated with similar task scenarios: such as sudden policy interpretation, industry financing event tracking, to provide adaptive strategy recommendations and tool combination optimization basis for subsequent tasks.

[0077] The above description is only the preferred embodiment of the present application and the explanation of the applied technical principles, and is not intended to limit the scope of the claimed invention, but only represents the preferred embodiment of the present application. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solutions formed by the specific combination of the above technical features, and also covers other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

Claims

1. A loop feedback scheduling collaboration method based on intelligent agents and cloud computing services, characterized in that: The following steps are involved: Step S1, memory recall: After receiving the task requirement, the agent accesses the memory library to obtain historical experience items, and performs memory matching and memory fusion locally based on the task requirement to generate memory recall information. The memory recall information is output to the cloud computing service for analysis. The cloud computing service includes a cloud-based large language model. Step S2, cloud-based task planning: Based on the memory recall information generated in step S1, the cloud-based large language model understands the task requirements, decomposes the tasks, and generates an executable description. The executable description includes the execution logic, dependencies, and execution goals of the task. The agent generates a preliminary task plan based on the executable description and calls the required tools from the tool set. Step S3, action observation: The agent performs the operation and observes the results, collecting real-time feedback data on the environment and the tools used, comparing the actual results with the expected goals, adjusting the task strategy, and selecting a more appropriate tool from the tool set for subsequent operations. The agent updates and loops through the task, repeatedly performing this step until the task is determined to be completed. Step S4, summary and reflection: The intelligent agent summarizes and reflects on the entire execution process of step S3, records and generates task execution trajectories, integrates them into experience entries and stores them in the memory bank in step S1. The experience entries include execution records, tool selection, operation steps, problems encountered and corresponding solutions.

2. The loop feedback scheduling collaboration method based on intelligent agents and cloud computing services according to claim 1 is characterized in that: The memory matching formula in step S1 is: Where q is the input vector of the current task; m i is the memory vector of the i-th historical experience entry in the memory bank; sim(q,m i ) is the cosine similarity. The historical experience item that is most similar to the current task q is selected by calculating the cosine similarity. When the cosine similarity is higher than the set threshold, it indicates that the historical experience item has a strong correlation with the current task and can be used for the memory fusion described in step S1.

3. The loop feedback scheduling collaborative method based on intelligent agents and cloud computing services according to claim 2 is characterized in that: The value range of the set threshold is 60%-80%.

4. The loop feedback scheduling collaborative method based on intelligent agents and cloud computing services according to claims 1 and 2 is characterized in that: The formula for memory fusion in step S1 is: S i =sim(q,m i )·λ i Where λ i is the logical level relevance weight, which is introduced based on the cosine similarity and represents the i-th historical experience entry m i The degree of logical relevance in the current task q; S i is the comprehensive relevance score.

5. The loop feedback scheduling collaboration method based on intelligent agents and cloud computing services according to claim 4 is characterized in that: The logical level relevance weight λ i The cloud-based large language model is adaptively adjusted according to the requirements of the task, and the value range is [0,1].

6. The loop feedback scheduling collaboration method based on intelligent agents and cloud computing services according to claim 1 is characterized in that: In step S2, the intention understanding and task decomposition of the task requirements are specifically as follows: let the current task be T, the intention understanding result vector of the task be I(T), and the memory recall information generated in step S1 be C. Then the formula for intention understanding and task decomposition is: I(T)=LLM intent (T) ST i =LLM decompose (I(T),C i ) Where, LLM intent (·) is the intent parsing capability function of the cloud-based large language model; C i is the context information used to decompose the i-th subtask; LLM decompose (·) is the task decomposition function of the large language model on the cloud; ST i is the i-th subtask.

7. The loop feedback scheduling collaboration method based on intelligent agents and cloud computing services according to claim 1 is characterized in that: The formula for generating the executable description in step S2 is: Exec_Desc′(T)=LLM generate_description ({ST1,ST φ ,...,ST N }) Where, {ST1,ST φ ,...,ST N } is the subtask sequence of the current task T; LLM generate_description (·) is the task summarization capability function of the cloud-based large language model; Exec_Desc′(T) is the executable description of the current task T.

8. The loop feedback scheduling collaboration method based on intelligent agents and cloud computing services according to claim 1 is characterized in that: The task is determined to be finally completed in step S3 when all subtasks are completed.

9. The loop feedback scheduling collaborative method based on intelligent agents and cloud computing services according to claim 1 is characterized in that: The formula for generating the task execution trajectory in step S4 is: Exec_Log(T)={[ST1,Tool1,O1],[ST φ ,Tool φ ,O0],...,[ST N ,Tool N ,O N ]} Where, ST1, ST φ ,...,ST N It is a subtask of the current task T; Tool1, Tool φ ,...,Tool N Tools used to complete this subtask; O1, O2, ..., O N is the observed execution result of the subtask; Exec_Log(T) is the execution track of the current task T.

Images