Intelligent agent system and control method thereof

Abstract

The invention discloses an intelligent agent system and a control method thereof, and relates to the technical field of natural language processing. The system comprises an execution layer, a planning layer and an auditing layer, the planning layer identifies and extracts innovation task meta-information, constructs a task tree to complete operations such as hierarchical clustering, and determines a task execution path; the execution layer calculates the semantic similarity between the content of the knowledge graph and the user innovation question, determines the optimal question and answer, extracts heuristic information, and decomposes the heuristic information into sub-questions to construct a dependency graph; generating an answer set based on the graph solving sub-problems, and integrating answers by using a genetic algorithm to obtain an overall solution; and the auditing layer performs multi-dimensional scoring on the scheme, and determines a structured scheme report and optimization suggestions for users to use according to a scoring result. The method has the capabilities of structured reasoning, problem recursive decomposition and scheme closed-loop optimization, can generate a multi-dimensional evaluation result, and efficiently realizes systematic modeling of a clear path for analogy heuristic information support problem solution.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to an intelligent agent system and its control method. Background Technology

[0002] Currently, with the continuous advancement of artificial intelligence technology in areas such as natural language processing, knowledge graphs, and reasoning modeling, an increasing number of research and applications are focusing on using AI to assist or replace humans in performing complex tasks. Among these, tasks, due to their high degree of goal uncertainty, open solution space, and high dependence on analogy and intuitive reasoning, have become one of the core challenges in the research of intelligent agent systems.

[0003] In existing technologies, although several tools have achieved certain results in solving innovative problems, some new tools, such as ChatGPT and Claude language models, have shown significant capabilities in question answering, summarizing, and code generation. Their foundation lies in pre-training on large-scale corpus data. In addition, Google DeepMind's Co-Scientist and Manus's intelligent agent system, which has multimodal capabilities, have good results in tool invocation and language generation.

[0004] However, the aforementioned existing technologies cannot perform structured reasoning on user innovation problems and lack systematic modeling that efficiently supports clear paths to solving innovation problems. Summary of the Invention

[0005] Therefore, it is necessary to provide an intelligent agent system and its control method to address the aforementioned technical problems.

[0006] The following technical solution is adopted in this specification: This specification provides an intelligent agent system, including: The system comprises an execution layer and a planning layer; the planning layer includes a task identification agent and a task planning agent; the execution layer includes an analogy agent, a decomposition agent, and an integration agent; and the review layer includes a review agent. The task identification intelligent agent is used to identify user questions and extract task meta-information of user questions; The task planning agent is used to construct a task tree based on the task meta-information, so as to perform task hierarchical clustering, node sorting and resource pre-allocation, and obtain the task execution path of the task meta-information. The analog agent is used to calculate the semantic similarity between the knowledge graph content and the user's question based on the task execution path of the task meta-information; determine the best question and its corresponding answer based on the similarity, and extract heuristic information from the best question and its corresponding answer; The decomposition agent is used to decompose heuristic information into multiple sub-problems and construct a dependency graph of the sub-problems; The integrated intelligent agent is used to solve each sub-problem based on the sub-problem dependency graph, generating a set of sub-problem answers; and to integrate the answers in the sub-problem answer set using a genetic algorithm to obtain the overall solution to the user problem.

[0007] Preferably, it further includes: The review layer includes: a review agent; The auditing agent is used to score the overall solution from multiple dimensions and obtain the scoring results; and based on the scoring results, it determines a structured solution report and optimization suggestions for solving the user's problem, for the user to refer to, reuse or further optimize.

[0008] Obtain user questions, identify user questions, and extract task meta-information of user questions; construct a task tree based on the task meta-information to perform task hierarchical clustering, node sorting, and resource pre-allocation to obtain the task execution path of the task meta-information; Based on the task execution path of the task metadata, the semantic similarity between the knowledge graph content and the user question is calculated; based on the similarity, the best question and its corresponding answer are determined, and heuristic information is extracted from the best question and its corresponding answer; the heuristic information is decomposed into multiple sub-questions, and a sub-question dependency graph is constructed; based on the sub-question dependency graph, each sub-question is solved to generate a sub-question answer set; a genetic algorithm is used to integrate the answers in the sub-question answer set to obtain the overall solution to the user question; The overall solution is scored from multiple dimensions to obtain the scoring results; and based on the scoring results, a structured solution report and optimization suggestions are determined for users to refer to, reuse or further optimize.

[0009] Preferably, the user problem is identified, and the task metadata of the user problem is extracted, specifically including: The user's question is input into the natural language processing model, and the task complexity of the user's question is calculated and the domain to which the user's question belongs is determined. The task complexity and the domain to which the task belongs are integrated into task metadata using GPT-4.

[0010] Preferably, a task tree is constructed based on the task metadata to perform hierarchical task clustering, node sorting, and resource pre-allocation, specifically including: The task metadata is processed, and the task is decomposed using Deepseek-R1. The decomposed task metadata is subjected to hierarchical clustering to obtain the task tree; Based on the task tree, resource pre-allocation and concurrent management are performed on the decomposed task metadata; The task execution path of the task metadata is obtained through API interface design after resource pre-allocation and concurrent management.

[0011] Preferably, the semantic similarity between the knowledge graph content and the user's question is calculated; based on the similarity, the best question and its corresponding answer are determined, and heuristic information is extracted from the best question and its corresponding answer, specifically including: Treat user questions as natural language questions; GPT-4 was used to perform analogy retrieval of natural language questions from knowledge graph content, resulting in multiple retrieval results; the retrieval results include the questions and their corresponding answers. Calculate the semantic similarity between each search result and the natural language question; The question with the highest semantic similarity is selected as the best question, and its corresponding answer is selected as the best answer. Use Deepseek-R1 to extract heuristics from the best questions and answers.

[0012] Preferably, the heuristic information is decomposed into multiple sub-problems, and a dependency graph of the sub-problems is constructed, specifically including: The heuristic information is recursively decomposed using Deepseek-R1 to obtain multiple subproblems; Multiple sub-problems are input into the knowledge graph, and GPT-4 is used to evaluate each sub-problem, including: If a subproblem can directly solve the user's problem, then the subproblem is retained. If a subproblem cannot directly solve the user's problem, then the subproblem is recursively decomposed using Deepseek-R1 again, and the judgment is made again, until the subproblem can directly solve the user's problem and is retained. Construct a dependency graph of subproblems based on the retained subproblems.

[0013] Preferably, each subproblem is solved to generate a set of answers to the subproblems, specifically including: We use GPT-4 to solve each subproblem and obtain the answer to each subproblem; and we calculate the semantic similarity between the answers to all subproblems. Based on the semantic similarity, answers exceeding a specified semantic similarity threshold are removed to obtain the answers to the sub-questions after compatibility testing; Based on the answers to the sub-questions after the compatibility check, construct a set of sub-question answers.

[0014] Preferably, the multi-dimensional scoring includes scores for the sex dimension, feasibility dimension, and consistency dimension.

[0015] Preferably, the structured solutions and optimization suggestions for the user problem are determined, specifically including: If the score is greater than or equal to the preset score threshold, a structured solution report and optimization suggestions for the user problem will be generated based on the overall solution. If the score is less than the preset score threshold, then based on the overall solution, feedback optimization suggestions are generated and fed back to the execution layer. The optimized overall solution is generated iteratively, and the overall solution after each iteration is scored from multiple dimensions until the score is greater than or equal to the preset score threshold. A structured solution report and optimization suggestions for the user problem are then generated.

[0016] This specification also provides a control method for the aforementioned intelligent agent system, including: Obtain user questions, identify user questions, and extract task meta-information of user questions; construct a task tree based on the task meta-information to perform task hierarchical clustering, node sorting, and resource pre-allocation to obtain the task execution path of the task meta-information; Based on the task execution path of the task metadata, the semantic similarity between the knowledge graph content and the user question is calculated; based on the similarity, the best question and its corresponding answer are determined, and heuristic information is extracted from the best question and its corresponding answer; the heuristic information is decomposed into multiple sub-questions, and a sub-question dependency graph is constructed; based on the sub-question dependency graph, each sub-question is solved to generate a sub-question answer set; a genetic algorithm is used to integrate the answers in the sub-question answer set to obtain the overall solution to the user question; The overall solution is scored from multiple dimensions to obtain the scoring results; and based on the scoring results, a structured solution report and optimization suggestions are determined for users to refer to, reuse or further optimize.

[0017] This specification also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned general problem-solving method based on combinatorial theory.

[0018] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned general problem-solving method based on combinatorial theory.

[0019] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: In the intelligent agent system and its control method provided in this specification, the user-input natural language problem is automatically structured into task meta-information and a task tree by using task identification and task planning agents in the planning layer, significantly improving the system's ability to express open-ended problems. By using analogy agents in the execution layer to perform structure mapping analogies, more transferable structural heuristic information can be accurately extracted compared to mechanisms that rely on pure semantic matching. By using decomposition agents in the execution layer to construct sub-problem dependency graphs, hierarchical recursive decomposition, concurrent scheduling, and resource planning of user problems are realized. By using integrated agents, the final solution is ensured to have evaluability and optimization feedback paths in multiple dimensions, including innovativeness, logical consistency, and practical feasibility.

[0020] In summary, this specification constructs a complete closed-loop intelligent agent system that integrates analogical reasoning, hierarchical decomposition, solution integration, and review feedback. Through its structured reasoning, recursive problem decomposition, and closed-loop solution optimization capabilities, it can generate multi-dimensional evaluation schemes that include innovation, logical consistency, and practical feasibility, thus efficiently realizing clear path modeling for innovative problems supported by analogical heuristic information. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0022] Figure 1 This is a schematic diagram of the system structure of an intelligent agent system provided in this specification; Figure 2 A flowchart illustrating a control method for an intelligent agent system provided in this specification; Figure 3 This is a schematic diagram of a computer device used to implement a control method for an intelligent agent system, as provided in this specification. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0024] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0025] Figure 1 This is a schematic diagram of the structure of an intelligent agent system described in this specification, specifically including: The system comprises an execution layer, a planning layer, and an approval layer; the planning layer includes a task identification agent and a task planning agent; the execution layer includes an analogy agent, a decomposition agent, and an integration agent; and the approval layer includes an approval agent. The task identification intelligent agent is used to identify user questions and extract task meta-information of user questions; The task planning agent is used to construct a task tree based on the task meta-information, so as to perform task hierarchical clustering, node sorting and resource pre-allocation, and obtain the task execution path of the task meta-information. The analog agent is used to calculate the semantic similarity between the knowledge graph content and the user's question based on the task execution path of the task meta-information; determine the best question and its corresponding answer based on the similarity, and extract heuristic information from the best question and its corresponding answer; The decomposition agent is used to decompose heuristic information into multiple sub-problems and construct a dependency graph of the sub-problems; The integrated intelligent agent is used to solve each sub-problem based on the sub-problem dependency graph, generating a set of answers to the sub-problems; and to integrate the answers in the set of answers to the sub-problems using a genetic algorithm to obtain the overall solution to the user problem. The auditing agent is used to score the overall solution from multiple dimensions and obtain the scoring results; and based on the scoring results, it determines a structured solution report and optimization suggestions for solving the user's problem, for the user to refer to, reuse or further optimize.

[0026] Optionally, the user problem is identified through the planning layer, and the task meta-information of the user problem is extracted. Specifically, this includes: inputting the user problem into the natural language processing model, calculating the task complexity of the user problem and determining the task domain of the user problem; and integrating the task complexity and task domain into task meta-information through GPT-4.

[0027] Specifically, for example, if a user inputs the question: "How can we improve users' motivation to continue learning on an online education platform?", the system first calls the natural language processing module to preprocess the question, identifying keywords such as "online education platform," "continuous learning," and "motivation." Then, the task identification module uses the Deepseek-R1 model to calculate the semantic complexity level of the question, initially determining that it involves multiple areas such as behavioral incentives, educational content optimization, and human-computer interaction design, classifying it as a high-complexity problem.

[0028] Next, the system performs task classification using the GPT-4 model. GPT-4 identifies the problem as being decomposed into the following primary task elements: ① User incentive mechanism design; ② Adaptability optimization of teaching content; ③ Construction of learning process visualization and feedback mechanisms; ④ Optimization of platform interface.

[0029] Optionally, a task tree is constructed based on the task metadata for hierarchical task clustering, node sorting, and resource pre-allocation. Specifically, this includes: processing the task metadata and decomposing the task using Deepseek-R1; performing hierarchical clustering on the decomposed task metadata to obtain the task tree; pre-allocating resources and managing concurrency based on the task tree; and designing the task execution path of the pre-allocated and concurrency-managed task metadata through an API interface.

[0030] Specifically, based on the first-level task metadata obtained above, the system then proceeds to the task tree construction stage. First, the system preprocesses the first-level task information, extracting key verbs (such as "design," "optimize," and "build") and target objects (such as "incentive mechanism" and "feedback system") from the task description through the task standardization module, forming a unified structured representation. Subsequently, the Deepseek-R1 model is used to recursively semantically decompose these task metadata. For example, "user incentive mechanism design" is further broken down into: ① constructing a behavior points system; ② introducing a social comparison mechanism (such as a leaderboard); ③ integrating achievement badges and reward mechanisms; "adaptive optimization of teaching content" is broken down into: ① learning behavior trajectory analysis; ② automatic content difficulty matching; ③ recommendation algorithm accuracy optimization, etc.

[0031] Based on the semantic similarity of the aforementioned multi-level tasks, the system constructs a hierarchical task tree and assigns metadata such as "complexity," "domain label," and "task dependencies" to each sub-task node. Building upon this, the system pre-allocates resources by invoking scheduling algorithms (such as topology sorting + load balancing strategies) based on task priority and computational resource consumption. Each task node specifies the required tools (such as a GPT-4 task executor, graph database retrieval tool, and evaluation module) and concurrent scheduling window. Finally, the system maps the aforementioned task metadata and their dependencies into a set of callable RESTful API interface paths.

[0032] Optionally, the semantic similarity between the knowledge graph content and the user's question is calculated through the execution layer; based on the similarity, the best question and its corresponding answer are determined, and heuristic information is extracted from the best question and its corresponding answer. Specifically, this includes: processing the user's question into a natural language question; using GPT-4 to perform analogy retrieval of the natural language question from the knowledge graph content to obtain multiple retrieval results; the retrieval results include the question and its corresponding answer; calculating the semantic similarity between each retrieval result and the natural language question; selecting the question with the highest semantic similarity as the best question, and its corresponding answer as the best answer; and using Deepseek-R1 to extract heuristic information from the best question and the best answer.

[0033] Specifically, to extract effective heuristics that can be used to solve the problem, the system integrates an analogical agent to perform a "structural similarity problem search" task. First, the system transforms the question "How to improve users' motivation for continuous learning on online education platforms?" into a semantic embedding vector using the Deepseek-R1 model and inputs it into a knowledge graph for similar question retrieval. The system matches a structurally similar historical question in the knowledge graph: "How to improve the long-term retention rate of fitness app users?" This question has a high structural mapping degree with the target question in terms of motivation, user behavior, and platform service mechanisms. After comparing the two questions, their semantic similarity reaches 0.91, and it is identified as the best analogy question.

[0034] The system further extracted solutions to this historical problem, including: ① introducing phased rewards; ② setting personalized goals; ③ visualizing feedback charts; and ④ guiding users to make public commitments (such as social media check-ins). These solutions were extracted as heuristics and used as input for subsequent problem decomposition phases.

[0035] Optionally, the heuristic information is decomposed into multiple sub-problems, and a sub-problem dependency graph is constructed. Specifically, this includes: recursively decomposing the heuristic information using Deepseek-R1 to obtain multiple sub-problems; inputting the multiple sub-problems into the knowledge graph, and using GPT-4 to judge each sub-problem, including: if the sub-problem can directly solve the user problem, then retain the sub-problem; if the sub-problem cannot directly solve the user problem, then recursively decompose the sub-problem again using Deepseek-R1 and judge it again, until the sub-problem can directly solve the user problem and is retained; and constructing a sub-problem dependency graph based on the retained sub-problems.

[0036] Specifically, the system inputs the above heuristic information into the Deepseek-R1 model, performs the first recursive problem decomposition, and obtains the following set of subproblems: Sub-question 1: How to design a phased learning reward system? Sub-question 2: How to achieve automatic generation and matching of personalized learning goals? Sub-question 3: How can visual feedback enhance the sense of accomplishment in the learning process? Sub-question 4: How to guide users to establish a commitment to open learning? Subsequently, the system calls GPT-4 to evaluate each sub-problem: For sub-problems 1 and 4, the system determines they can be directly designed and executed at the platform level; for sub-problems 2 and 3, due to their involvement in individual modeling and visual cognitive mechanisms, the system determines they cannot be directly answered and require further decomposition. Therefore, the system again calls Deepseek-R1 to decompose sub-problem 2, generating the following secondary sub-problems:

[0037] Sub-question 2.1: How to build a user learning preference profile based on learning history data? Sub-question 2.2: How to locate content modules in a knowledge graph that match the learning objectives? Finally, the system retains all the above subproblems and constructs a subproblem dependency graph with the main problem as the root node.

[0038] Optionally, each subproblem is solved to generate a set of subproblem answers. Specifically, this includes: solving each subproblem using GPT-4 to obtain the answer to each subproblem; calculating the semantic similarity between the answers to all subproblems; removing answers with a semantic similarity greater than a specified threshold based on the semantic similarity to obtain the subproblem answers after compatibility testing; and constructing the subproblem answer set based on the subproblem answers after compatibility testing.

[0039] Specifically, for the set of subproblems constructed in the previous stage, the system calls GPT-4 to solve each subproblem individually: Sub-question 1: How to design a phased learning reward system? GPT-4 provides the answer: Build a points-based growth mechanism of "short-term incentives + long-term goals". Points, badges or redemption codes are automatically awarded for each learning milestone achieved (such as logging in for 7 consecutive days or completing a quiz).

[0040] Sub-question 2.1: How to build a user learning preference profile based on learning history data? Answer: Collect user behavior logs (browsing frequency, completion time, etc.), combine graph neural networks to dynamically model user preferences, and generate interpretable profile labels.

[0041] Sub-question 2.2: How to locate content modules that match learning objectives in a knowledge graph? Answer: Perform semantic matching between the user profile embedding vector and the course content node vector, and filter content based on cosine similarity to select recommended modules with a similarity greater than 0.85.

[0042] Sub-question 3: How to enhance the sense of accomplishment in the learning process through visual feedback? GPT-4 generated two candidate answers: ① Construct a learning "journey map" to show the learning stages; ② Display a "growth tree" or "radar chart" in real time to track skill improvement.

[0043] Sub-question 4: How to guide users to establish a public learning commitment mechanism? Answer: Introduce the "social commitment wall" function, allowing users to post learning goals and generate shareable images and text, supporting a check-in and supervision mechanism on social media.

[0044] The system calculates the semantic compatibility among all the above answers. Analysis reveals that the two alternatives for sub-question 3 have a slight stylistic conflict (the journey map is linear, while the radar chart is dimensionally expanded), and their semantic compatibility score is only 0.64; the system automatically selects the "journey map" as the retained option and discards the radar chart.

[0045] The final set of answers to the sub-problems is as follows: ① Points-based growth mechanism; ② User profile generation methods; ③ Recommended content matching strategy; ④ Learning journey map feedback; ⑤ Social commitment wall mechanism.

[0046] This subset of answers will serve as input for the next step of optimizing the "integrated agent" scheme structure.

[0047] Optionally, the multi-dimensional rating includes ratings for the sex dimension, feasibility dimension, and consistency dimension.

[0048] Optionally, based on the scoring results, a structured solution for the user problem is determined, specifically including: if the scoring result is greater than or equal to a preset scoring threshold, a structured solution report and optimization suggestions for the user problem are generated based on the overall solution; if the scoring result is less than the preset scoring threshold, feedback optimization suggestions are generated based on the overall solution and fed back to the execution layer, and an optimized overall solution is iteratively generated. The optimized overall solution is scored in multiple dimensions for each iteration until the scoring result is greater than or equal to the preset scoring threshold, and a structured solution report and optimization suggestions for the user problem are generated.

[0049] Specifically, after the sub-question answer set is integrated, the system inputs the complete preliminary solution into the review agent, and the Claude 3.7 model performs a three-dimensional scoring, including: ① Sexual dimension: Whether novel mechanisms are introduced (such as points-based growth + social commitment wall); ② Feasibility dimension: Does the platform already have the technical foundation to implement the various recommendations (such as log collection and graph visualization)? ③ Consistency dimension: Whether the logic of each sub-module is consistent, the expression style is unified, and there are no conflicts or redundancies.

[0050] In this example, the preliminary integration scheme received the following scores: ① Feasibility: 8.7 / 10; ② Consistency: 6.2 / 10; ③ Consistency: 7.0 / 10; Overall score: 7.3 / 10, slightly lower than the preset threshold (7.5).

[0051] Based on the scoring results, the system automatically generated the following optimization suggestion report: The second submodule, "Personalized Recommendation," and the fifth submodule, "Social Commitment," lacked workflow linkage. It was suggested that a "Recommendation Target Release Mechanism" be embedded to bind the system's recommended content to user commitment tasks. Subsequently, the system fed back the optimization suggestions to the "Integrated Intelligent Agent" execution layer. GPT-4 received the suggestions and adjusted the solution process, embedding the recommended content into the commitment release process in the form of diagram cards, forming a linked task package.

[0052] The new integration solution was input into the review agent again and re-evaluated in three dimensions: ① Feasibility: 9.0; ② Consistency: 7.5; ③ Consistency: 8.0; Overall score: 8.2, exceeding the preset threshold.

[0053] The system ultimately confirmed that the solution was a structured output solution and generated a structured solution report, which included module division, functional flowcharts, required resources and API call tables, as well as optimization suggestions and iteration logs, supporting further deployment or submission for user review.

[0054] based on Figure 1 The general problem-solving method based on combinatorial theory shown can be used as a scientific research tool to help universities or research institutions generate research hypotheses, propose new questions, and evaluate multiple solutions, saving mental labor; it can be used as a product design platform to quickly analyze user needs, generate creative solutions, and evaluate their feasibility in industrial design and services; it can be used as a general task-solving agent to connect with multiple domain models and knowledge bases through open interfaces to realize cross-industry AI Co-pilot services, and is expected to lead to SaaS products.

[0055] According to preliminary tests, AI Co-Innovator's results in hypothesis generation tasks achieved an average expert score 25% higher than the control group system in terms of "novelty" and "structural rigor," demonstrating significant commercial value.

[0056] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0057] The above are general problem-solving methods based on combinatorial theory provided by one or more embodiments of this specification. Based on the same idea, this specification also provides a corresponding control method for an intelligent agent system, such as... Figure 2 As shown.

[0058] Figure 2 This specification provides a flowchart of a control method for an intelligent agent system, which specifically includes the following steps: S101: Obtain user questions, identify user questions, and extract task meta-information of user questions; construct a task tree based on the task meta-information to perform task hierarchical clustering, node sorting, and resource pre-allocation to obtain the task execution path of the task meta-information. S102: Based on the task execution path of the task metadata, calculate the semantic similarity between the knowledge graph content and the user question; based on the similarity, determine the best question and its corresponding answer, and extract heuristic information from the best question and its corresponding answer; decompose the heuristic information into multiple sub-questions, and construct a sub-question dependency graph; based on the sub-question dependency graph, solve each sub-question to generate a sub-question answer set; use a genetic algorithm to integrate the answers in the sub-question answer set to obtain the overall solution to the user question; S103: Evaluate the overall solution from multiple dimensions to obtain the evaluation results; and based on the evaluation results, determine the structured solution report and optimization suggestions for solving the user's problem, for the user to refer to, reuse or further optimize.

[0059] Specifically, the user inputs a natural language question Q; the planning layer identifies the task domain and complexity using Deepseek-R1 and BERT models, constructing a task meta-information tuple T(Q); the planning agent generates a task tree G(Q) based on task clustering and complexity evaluation results, and calls topological sorting to construct the execution path; analogous to the agent identifying structurally similar questions QS from a knowledge graph, ViMimic is used for structure mapping to extract heuristic information I(Q). S The decomposition agent receives I, and the recursive decomposition problem Q is {q1,q2,...,q}. n}, and construct the subproblem dependency graph D(Q); for each q i The solution is obtained by calling GPT-4, generating a solution set A = {a1, a2, ..., a...} n The integrated agent uses semantic similarity to check for conflicts or repetitions between solutions in A, and uses a genetic algorithm to optimize and combine them into an overall solution A*. The review agent evaluates and scores A* by scoring S(A*). If the score is lower than the threshold θ, it provides feedback to the analog agent (re-analyzing) for iterative solution. The approved structured solution A*_final is output for users to refer to, reuse or further optimize.

[0060] For specific limitations on an intelligent agent system, please refer to the limitations on general problem-solving methods based on combinatorial theory mentioned above, which will not be repeated here. Each module in the aforementioned general problem-solving intelligent agent system based on combinatorial theory can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0061] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 2 A control method for an intelligent agent system is provided.

[0062] This instruction manual also provides Figure 3 The schematic diagram of the computer device shown is as follows: Figure 3 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 2 A control method for an intelligent agent system is provided.

[0063] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this specification can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

Claims

1. An intelligent agent system, characterized in that, include: Execution layer and planning layer; The planning layer includes a task identification agent and a task planning agent; the execution layer includes an analogy agent, a decomposition agent, and an integration agent; the review layer includes a review agent. The task identification intelligent agent is used to identify user questions and extract task meta-information of user questions; The task planning agent is used to construct a task tree based on the task meta-information, so as to perform task hierarchical clustering, node sorting and resource pre-allocation, and obtain the task execution path of the task meta-information. The analog agent is used to calculate the semantic similarity between the knowledge graph content and the user's question based on the task execution path of the task meta-information; determine the best question and its corresponding answer based on the similarity, and extract heuristic information from the best question and its corresponding answer; The decomposition agent is used to decompose heuristic information into multiple sub-problems and construct a dependency graph of the sub-problems; The integrated intelligent agent is used to solve each sub-problem based on the sub-problem dependency graph, generating a set of sub-problem answers; and to integrate the answers in the sub-problem answer set using a genetic algorithm to obtain the overall solution to the user problem.

2. The intelligent agent system as described in claim 1, characterized in that, Also includes: Review layer; The review layer includes: a review intelligent agent; The auditing agent is used to score the overall solution from multiple dimensions and obtain the scoring results; and based on the scoring results, it determines a structured solution report and optimization suggestions for solving the user's problem, for the user to refer to, reuse or further optimize.

3. The intelligent agent system as described in claim 1, characterized in that, The process of identifying user questions and extracting task metadata about those questions specifically includes: The user's question is input into the natural language processing model, and the task complexity of the user's question is calculated and the domain to which the user's question belongs is determined. The task complexity and the domain to which the task belongs are integrated into task metadata using GPT-4.

4. The intelligent agent system as described in claim 1, characterized in that, The step of constructing a task tree based on the task metadata, performing task hierarchical clustering, node sorting, and resource pre-allocation to obtain the task execution path of the task metadata, specifically includes: The task metadata is processed, and the task is decomposed using Deepseek-R1. The decomposed task metadata is subjected to hierarchical clustering to obtain the task tree; Based on the task tree, resource pre-allocation and concurrent management are performed on the decomposed task metadata; The task execution path of the task metadata is obtained through API interface design after resource pre-allocation and concurrent management.

5. The intelligent agent system as described in claim 1, characterized in that, The process involves calculating the semantic similarity between the knowledge graph content and the user's question; based on the similarity, determining the best question and its corresponding answer, and extracting heuristic information from the best question and its corresponding answer, specifically including: Treat user questions as natural language questions; GPT-4 was used to perform analogy retrieval of natural language questions from knowledge graph content, resulting in multiple retrieval results; the retrieval results include the questions and their corresponding answers. Calculate the semantic similarity between each search result and the natural language question; The question with the highest semantic similarity is selected as the best question, and its corresponding answer is selected as the best answer. Use Deepseek-R1 to extract heuristics from the best questions and answers.

6. The intelligent agent system as described in claim 1, characterized in that, The process of decomposing heuristic information into multiple sub-problems and constructing a dependency graph of these sub-problems specifically includes: The heuristic information is recursively decomposed using Deepseek-R1 to obtain multiple subproblems; Multiple sub-problems are input into the knowledge graph, and GPT-4 is used to evaluate each sub-problem, including: If a subproblem can directly solve the user's problem, then the subproblem is retained. If a subproblem cannot directly solve the user's problem, then the subproblem is recursively decomposed using Deepseek-R1 again, and the judgment is made again, until the subproblem can directly solve the user's problem and is retained. Construct a dependency graph of subproblems based on the retained subproblems.

7. The intelligent agent system as described in claim 1, characterized in that, The process of solving each subproblem and generating a set of answers for each subproblem specifically includes: We use GPT-4 to solve each subproblem and obtain the answer to each subproblem; and we calculate the semantic similarity between the answers to all subproblems. Based on the semantic similarity, answers exceeding a specified semantic similarity threshold are removed to obtain the answers to the sub-questions after compatibility testing; Based on the answers to the sub-questions after the compatibility check, construct a set of sub-question answers.

8. The intelligent agent system as described in claim 2, characterized in that, The multi-dimensional scoring includes scores for the sex dimension, feasibility dimension, and consistency dimension.

9. The intelligent agent system as described in claim 2, characterized in that, The process of determining structured solutions and optimization suggestions for user problems based on the scoring results specifically includes: If the score is greater than or equal to the preset score threshold, a structured solution report and optimization suggestions for the user problem will be generated based on the overall solution. If the score is less than the preset score threshold, then based on the overall solution, feedback optimization suggestions are generated and fed back to the execution layer. The optimized overall solution is generated iteratively, and the overall solution after each iteration is scored from multiple dimensions until the score is greater than or equal to the preset score threshold. A structured solution report and optimization suggestions for the user problem are then generated.

10. A control method for a general intelligent agent system according to any one of claims 1 to 9, characterized in that, include: Obtain user questions, identify user questions, and extract task metadata about user questions; Based on the task metadata, a task tree is constructed to perform task hierarchical clustering, node sorting, and resource pre-allocation, thereby obtaining the task execution path of the task metadata. Based on the task execution path of the task metadata, calculate the semantic similarity between the knowledge graph content and the user's question; based on the similarity, determine the best question and its corresponding answer, and extract heuristic information from the best question and its corresponding answer; decompose the heuristic information into multiple sub-questions, and construct a sub-question dependency graph; Based on the sub-problem dependency graph, each sub-problem is solved to generate a set of answers for the sub-problems; a genetic algorithm is then used to integrate the answers from the sub-problem answer sets to obtain the overall solution to the user's problem. The overall solution is scored from multiple dimensions to obtain the scoring results; and based on the scoring results, a structured solution report and optimization suggestions are determined for users to refer to, reuse or further optimize.

Images