A multi-agent collaborative system and strategy method suitable for industrial digitalization

Abstract

The present invention discloses a multi-agent collaborative system and strategy method suitable for industrial digitalization. The system includes: a large model system for realizing central management functions; a tool library for providing basic functions; a functional model library for providing analytical operations; according to the input task, the large model system obtains and understands the user request corresponding to the task, and decomposes the task into multiple goals and several subtasks corresponding to each goal, dispatches the tool corresponding to each subtask from the tool library, and dispatches the functional model corresponding to each subtask from the functional model library, completes the combined calculation or decision of data and models, and forms a strategy plan corresponding to the input task; the system solves the problem that the workload generated by the data middle station in solving the industrial business system collaboration and data value mining is large and cannot be coordinated with business adjustments, and at the same time, it can also reduce the dependence on engineers' data information processing capabilities to achieve the collaboration of upstream and downstream of the industrial chain.

Description

Technical Field

[0001] The present invention relates to the field of multi-agent collaboration, and more specifically, to a multi-agent collaboration system and strategy method suitable for industrial digitalization. Background Art

[0002] With the development of information technology and the development of industrial digitalization, the construction of information systems and the establishment of application scenarios are becoming more and more perfect at the industrial end. How to mine data value from a complete business system and how to move from business interconnection to data-driven decision-making has become a new development stage of industrial digitalization. At present, the data center is mainly responsible for data aggregation, data analysis, data value mining exploration, and integration and interconnection of various business systems. However, there are three problems in solving the problems of industrial business system collaboration and data value mining with the data center: 1. The workload is large. The data generated in various links of the industry is massive. The processing of information depends on the experience of engineers to build a large number of data analysis models. Data acquisition, cleaning, feature processing, modeling, etc. take a lot of time; 2. The positioning of the data center often focuses on data processing, and the adjustment of workload and business cannot be effectively coordinated. For example, the OA system often formulates the approval flow of production report change style, authority, etc., but the generation of production reports is completed in the ERP system. The data center cannot coordinate business flow and data flow; 3. The collaboration of upstream and downstream of the industrial chain depends on the ability of engineers to process data and information, and cannot actively mine data value.

[0003] Therefore, in order to solve the above problems, the present invention proposes a multi-agent collaborative system and strategy method suitable for industrial digitalization, which is used to solve the problem that the data middle platform has a large workload and cannot coordinate with business adjustments when solving industrial business system collaboration and data value mining. At the same time, it can also reduce the dependence on engineers' data information processing capabilities to achieve collaboration between upstream and downstream of the industrial chain. Summary of the invention

[0004] The present invention aims to overcome at least one defect (shortcoming) of the above-mentioned prior art and provide a multi-agent collaborative system and strategy method suitable for industrial digitalization, which is used to solve the problem that the data middle platform has a large workload and cannot coordinate with business adjustments when solving industrial business system collaboration and data value mining. At the same time, it can also reduce the dependence on engineers' data information processing capabilities to achieve collaboration between upstream and downstream of the industrial chain.

[0005] The technical solution adopted by the present invention is a multi-agent collaborative system suitable for industrial digitalization, comprising:

[0006] A large model system for implementing central management functions;

[0007] A tool library to provide basic functions;

[0008] A functional model library for providing analytical operations;

[0009] According to the input task, the large model system obtains and understands the user request corresponding to the task, decomposes the task into multiple goals and several subtasks corresponding to each goal, dispatches the tool corresponding to each subtask from the tool library, dispatches the functional model corresponding to each subtask from the functional model library, completes the combined calculation or decision of data and models, and forms a strategy plan corresponding to the input task;

[0010] Through this system, the text entered by the user can be divided into tasks, and then each business process can be divided into intelligent entities. Through a large model system with powerful central management functions, the collaborative collaboration of each intelligent entity can be achieved, reducing workload and improving work efficiency.

[0011] Preferably, it also includes a domain database related to the domain involved in the task and a memory module for implementing long and short-term memory;

[0012] The large model system calls relevant data in the domain database according to the input task, so that the large model system can further understand the domain knowledge related to the task. The memory module is used to record previous strategy plans to provide a reference basis for the large model system to understand the context of user requests and the planning of each subtask.

[0013] By actively mining data, obtaining relevant data from the domain database corresponding to the task and previously recorded strategy plans, the large model can better understand the domain knowledge related to the task, thereby breaking down the task into more matching sub-goals, thereby reducing dependence on data analysis engineers, further reducing workload and improving work efficiency.

[0014] Preferably, the multi-agent collaborative system suitable for industrial digitalization also includes an industrial business system and a comprehensive analysis and control system; the industrial business system is used to provide the large model system with real-time operational data of the industry; the comprehensive analysis and control system is used to provide the large model system with periodic analysis data of the industry. In order to further fine-tune and achieve precision in task processing, an industrial business system and a comprehensive analysis and management system are also provided in this system, so that the general large model can be matched to domain knowledge. The large model system can be refined to specific industrial parks when performing goal and task decomposition, and obtain the recent analysis data of the corresponding industrial park, so that the output strategy plan has a higher matching degree and more business value.

[0015] Further preferably, a manual feedback module is also included;

[0016] The manual feedback module is used to provide the user with prompt word input feedback. The large model system further decomposes the task into multiple goals and several subtasks corresponding to each goal according to the prompt word feedback from the user. By providing manual feedback to the user for prompt word input, the completeness of the task described by the user can be further improved, so that the large model system can further understand the task details, thereby achieving more accurate task strategy planning.

[0017] More preferably, the manual feedback module includes:

[0018] A knowledge base of all text generation task elements needed to complete a task;

[0019] A vector database of all the text generation task descriptions mentioned above;

[0020] A NOSQL database storing a text generation task description and text generation knowledge based on the text generation task description; the specific process of the manual feedback module for providing feedback to the user on the prompt word input is as follows:

[0021] S41: The large model system obtains and understands the user request corresponding to the task and forms an initial prompt;

[0022] S42: vectorizing the initial Prompt text, and matching the text generation task description that best meets the task requirements in the vector database of text generation task descriptions;

[0023] S43: according to the text generation description, query the text generation knowledge that completes this task in the NOSQL database of the text generation task knowledge base;

[0024] S44: constructing a system-level prompt, wherein the system-level prompt specifically refers to determining whether the user has fully described the text generation task when the large model system receives knowledge about the text generation task;

[0025] S45: Based on the text generation knowledge obtained in S43 and the system-level prompt in S44, it is determined whether it is necessary to provide the user with a prompt word input;

[0026] S46: if necessary, iterate the text of the prompt words fed back by the user to generate a task description;

[0027] S47: Repeat S45 and S46 until the large model confirms that the user has completely described his text generation task, and then generates text based on the final text generation task description.

[0028] Through the above steps, the task description is updated and iterated, which ultimately improves the completeness of the task text description input by the user into the system, so that the output strategy plan has a higher matching degree and more business value.

[0029] Preferably, the multi-agent collaborative system suitable for industrial digitalization also includes a capability judgment module, which is used to judge the matching degree between the tools dispatched from the tool library and the functional models dispatched from the functional model library by the large model system and the strategy scheme corresponding to the input task. By judging the matching degree, the effect of the task completed by the system is evaluated. If the matching degree is poor, the system will perform update learning and further self-optimization.

[0030] Preferably, the system further includes a tool production module. When the large model system determines through the capability judgment module that the selected tool cannot achieve the expected matching degree, the tool production module is called to produce a tool with a higher matching degree. By producing a tool with a higher matching degree to perform the required task, self-update is achieved, and the versatility and intelligence of the entire large model system are improved.

[0031] Preferably, a multi-agent collaborative strategy method suitable for industrial digitalization includes the following steps:

[0032] S1: Task planning: Use the language model to decompose the input task into multiple sub-goals;

[0033] S2: Task parsing: Understand the user requests for multiple sub-goals based on the domain database related to the domain involved in the task and the strategy solutions previously recorded in the memory module, thereby generating multiple corresponding sub-tasks;

[0034] S3: Tool and model selection: For each subtask, select the tool and functional model to implement the subtask;

[0035] S4: Task execution: According to the selected tools and functional models, the corresponding subtasks are executed to achieve their respective sub-goals;

[0036] S5: Strategy generation: integrating multiple sub-goals to form an overall strategy plan;

[0037] Through the above steps, the task input by the user can be intelligently decomposed into multiple sub-goals, and the corresponding sub-tasks can be generated and executed, thereby reducing the dependence on data analysis engineers, further reducing the overall workload and improving work efficiency.

[0038] Preferably, step S1 also includes performing manual feedback updates. If the language model receives knowledge about the text generation task and determines that the user has not fully described the task of the text, it is necessary to provide the user with prompt word input until the text input by the user can fully describe the generated task and decompose the target into multiple sub-targets. By performing manual feedback updates, the task text description of the language model input by the user can be made more complete, thereby ensuring the accuracy of the final result generation.

[0039] Preferably, step S2 also includes fine-tuning learning and expert optimization, by acquiring real-time operational data of the industrial park and periodic data of the industry to perform fine-tuning learning, and introducing experts to assist in optimization, so as to achieve more accurate task analysis.

[0040] Preferably, the step S5 also includes capability judgment and tool production. If it is judged that the tools dispatched by the language big model from the tool library and the functional models dispatched from the functional model library are poorly matched with the strategy solutions corresponding to the input tasks, tool generation is performed, and the language big model produces new tools with higher matching degree by itself, and then step S4 is re-executed to execute multiple subtasks, thereby further realizing the intelligence of the language big model and improving its versatility.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] The present invention proposes a multi-agent collaborative system and strategy method suitable for industrial digitalization, thereby improving the collaboration between various intelligent agents in the data center when solving industrial business. By performing intelligent self-data mining and analysis, the data processing capability is improved, the dependence on data analysts is reduced, and the workload is greatly reduced, thereby improving work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a system structure diagram of the present invention.

[0044] Figure 2 It is a system flow chart of the present invention.

[0045] Figure 3 It is a manual feedback flow chart of the present invention.

[0046] Figure 4 Schematic diagram of the strategy method of the present invention.

[0047] Figure 5 Schematic diagram of fine-tuning learning and expert optimization of the present invention. DETAILED DESCRIPTION

[0048] The drawings of the present invention are only for illustrative purposes and should not be construed as limiting the present invention. In order to better illustrate the following embodiments, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; it is understandable to those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0049] Example 1

[0050] like Figure 1 As shown, this embodiment provides a multi-agent collaborative system suitable for industrial digitalization, including:

[0051] A large model system for implementing central management functions;

[0052] A tool library to provide basic functions;

[0053] A functional model library for providing analytical operations;

[0054] According to the input task, the large model system obtains and understands the user request corresponding to the task, decomposes the task into multiple goals and several subtasks corresponding to each goal, dispatches the tool corresponding to each subtask from the tool library, dispatches the functional model corresponding to each subtask from the functional model library, completes the combined calculation or decision of data and models, and forms a strategy plan corresponding to the input task;

[0055] Through this system, the text entered by the user can be divided into tasks, and then each business process can be divided into intelligent entities. Through a large model system with powerful central management functions, the collaborative collaboration of each intelligent entity can be achieved, reducing workload and improving work efficiency.

[0056] Preferably, if Figure 1 As shown, the system also includes a domain database related to the field involved in the task and a memory module for realizing long and short-term memory; the domain database is stored horizontally, and contains data on the fields of various industries, including transportation management systems, construction standards, investment information and development plans, etc., to provide more data references for task analysis in the large model system; at the same time, the memory module is stored vertically, and when analyzing the task, the memory module can be called to find out the strategy corresponding to the task before, and use it as a reference to further improve the final task analysis result.

[0057] The large model system calls relevant data in the domain database according to the input task, so that the large model system can further understand the domain knowledge related to the task. The memory module is used to record previous strategy plans to provide a reference basis for the large model system to understand the context of user requests and the planning of each subtask.

[0058] By actively mining data, obtaining relevant data from the domain database corresponding to the task and previously recorded strategy plans, the large model can better understand the domain knowledge related to the task, thereby breaking down the task into more matching sub-goals, thereby reducing dependence on data analysis engineers, further reducing workload and improving work efficiency.

[0059] Preferably, if Figure 1 As shown, the multi-agent collaborative system suitable for industrial digitalization also includes an industrial business system and a comprehensive analysis and control system. The industrial business system contains information on energy consumption, property, office, security, etc. of each industry, and the comprehensive analysis and control system also conducts comprehensive analysis and control on the information stored in the industrial business system, such as energy conservation and emission reduction, operation process analysis, investment promotion and property analysis.

[0060] Furthermore, the industrial business system provides the big model system with real-time operational data of the industry, and the comprehensive analysis and control system provides the big model system with periodic analysis data of the industry. Through further fine-tuning to achieve precision in task processing, the general big model can match knowledge in various fields. When breaking down goals and tasks, the big model system can be refined to specific industrial parks, and obtain the recent analysis data of the corresponding industrial parks, so that the output strategy solutions have a higher matching degree and more business value.

[0061] More preferably, if Figure 2 As shown, a manual feedback module is also included;

[0062] The manual feedback module is used to provide the user with prompt word input feedback. The large model system further decomposes the task into multiple goals and several subtasks corresponding to each goal according to the prompt word feedback from the user. By providing manual feedback to the user for prompt word input, the completeness of the task described by the user can be further improved, so that the large model system can further understand the task details, thereby achieving more accurate task strategy planning.

[0063] More preferably, the manual feedback module includes:

[0064] A knowledge base of all text generation task elements needed to complete a task;

[0065] A vector database of all the text generation task descriptions mentioned above;

[0066] A NOSQL database storing a text generation task description and text generation knowledge based on the text generation task description; Figure 3 As shown, the specific process of the manual feedback module for providing feedback to the user on the prompt word input is as follows:

[0067] S41: The large model system obtains and understands the user request corresponding to the task and forms an initial prompt;

[0068] S42: vectorizing the initial Prompt text, and matching the text generation task description that best meets the task requirements in the vector database of text generation task descriptions;

[0069] S43: according to the text generation description, query the text generation knowledge that completes this task in the NOSQL database of the text generation task knowledge base;

[0070] S44: constructing a system-level prompt, wherein the system-level prompt specifically refers to determining whether the user has fully described the text generation task when the large model system receives knowledge about the text generation task;

[0071] S45: Based on the text generation knowledge obtained in S43 and the system-level prompt in S44, it is determined whether it is necessary to provide the user with a prompt word input;

[0072] S46: if necessary, iterate the text of the prompt words fed back by the user to generate a task description;

[0073] S47: Repeat S45 and S46 until the large model confirms that the user has completely described his text generation task, and then generates text based on the final text generation task description.

[0074] For example, the task description text entered by the user into the system is "Pig Market Quotes".

[0075] In step S41, the large model system first understands the acquired "live pig market quotation" and forms an initial prompt;

[0076] In step S42, the initial Prompt is vectorized, and the text generation task description that best meets the task requirements is matched in the vector database of text generation task descriptions as "live pigs, live pig market quotes";

[0077] In step S43, according to the text generation description, the text generation knowledge to complete this task is queried in the NOSQL database of the text generation task knowledge base, which is "pig production area or brand, pig breed, market price, quotation provider, quotation release time, etc.";

[0078] In step S44, a system-level prompt is constructed to determine whether the user has fully described his text generation task;

[0079] In step S45, according to the task description "live pigs, live pig market quotation" generated in S42, and the text-generated knowledge "live pig origin or brand, live pig breed, market price, quotation provider, quotation release time, etc." obtained in S43, combined with the system-level prompt in S44, it is determined that the prompt word input needs to be fed back to the user;

[0080] In step S46, according to the prompt words fed back by the user, such as adding prompt words such as specific pig production area or brand, pig breed, quotation release time, etc., iterate its text to generate task description;

[0081] In step S47, S45 and S46 are repeatedly executed until the large model confirms that the user has fully described the text generation task, such as: the origin or brand of live pigs is Changsha, Hunan, the live pig breed is three-way foreign pigs, the quotation provider is Changsha market, Hunan, and the release time is September 2023, etc. Task prompt words, but not limited to the above text input, prompt words can be input according to actual conditions, and finally text is generated based on the final text generation task description.

[0082] Through the above steps, the task description is updated and iterated, which ultimately improves the completeness of the task text description input by the user into the system, so that the output strategy plan has a higher matching degree and more business value.

[0083] Preferably, if Figure 2 As shown, the multi-agent collaborative system suitable for industrial digitalization also includes a capability judgment module, which is used to judge the matching degree between the tools dispatched from the tool library and the functional models dispatched from the functional model library by the large model system and the strategy scheme corresponding to the input task. By judging the matching degree, the effect of the task completed by the system is evaluated. It has learning ability, iterative update ability, etc. If the matching degree between the tool dispatched from the tool library and the strategy scheme corresponding to the input task is poor, the large model system will update and learn, further perform self-optimization, generate tools that meet the task requirements and output them.

[0084] Further preferably, it also includes a tool production module. When the large model system determines through the capability judgment module that the selected tool cannot achieve the expected matching degree, the tool production module is called to produce a tool with a higher matching degree. In this embodiment, the large model system can act as a tool manufacturer to create tools for a given task. The tools are usually implemented in the form of python utility functions. By producing tools with a higher matching degree by themselves to perform the required tasks, self-update and iteration are achieved. At the same time, the large model system can also act as a tool user, using the tools created by the tool maker to solve problems. At the same time, the tool can also be expanded, with the help of frameworks such as Langchain, HunggingGPT, ModelScopeGPT, AutoGPT, etc., extension tools such as search, expert models, API, knowledge graphs, etc., to improve the versatility, practicality and intelligence of the entire large model system.

[0085] Preferably, if Figure 4 As shown, a multi-agent collaborative strategy method suitable for industrial digitalization includes the following steps:

[0086] S1: Task planning: Use the language model to decompose the input task into multiple sub-goals;

[0087] S2: Task parsing: Understand the user requests for multiple sub-goals based on the domain database related to the domain involved in the task and the strategy solutions previously recorded in the memory module, thereby generating multiple corresponding sub-tasks;

[0088] S3: Tool and model selection: For each subtask, select the tool and functional model to implement the subtask;

[0089] S4: Task execution: According to the selected tools and functional models, the corresponding subtasks are executed to achieve their respective sub-goals;

[0090] S5: Strategy generation: integrating multiple sub-goals to form an overall strategy plan;

[0091] Through the above steps, the task input by the user can be intelligently decomposed into multiple sub-goals, and the corresponding sub-tasks can be generated and executed, thereby reducing the dependence on data analysis engineers, further reducing the overall workload and improving work efficiency.

[0092] Preferably, step S1 also includes performing manual feedback updates. If the language model receives knowledge about the text generation task and determines that the user has not fully described the task of the text, it is necessary to provide the user with prompt word input until the text input by the user can fully describe the generated task and decompose the target into multiple sub-targets. By performing manual feedback updates, the task text description of the language model input by the user can be made more complete, thereby ensuring the accuracy of the final result generation.

[0093] Preferably, if Figure 5 As shown, step S2 also includes fine-tuning learning and expert optimization. Fine-tuning learning is performed by acquiring real-time operation data of the industrial park and periodic data of the industry, and experts are introduced to assist in optimization, so as to achieve more accurate task analysis.

[0094] For example, in this embodiment, for the energy management process, our task is to provide energy conservation and emission reduction suggestions according to energy policies, analyze energy consumption, and optimize and adjust the energy supply of the park. This task description is submitted to the large language model, which learns the specific details of the energy management SOP, energy analysis indicator system, and energy management standards and specifications through fine-tuning learning. At the same time, it uses expert optimization to assist in finding the key components of the energy management process, obtains detailed information on each part of the energy management process, and conducts a comprehensive analysis of energy process management based on the information provided.

[0095] Then, based on the information related to the energy management process obtained above, the task is broken down into four subtasks:

[0096] Step 1: Search the entire network for the latest energy policies and park energy management systems to generate an energy policy report.

[0097] Step 2: Link to the energy management system to analyze the energy consumption trend of the park in the past five years, link to the property relationship system, associate the property and park office enterprises to analyze the energy consumption, and generate a report;

[0098] Step 3: Link the park energy management expert model, industry energy conservation and emission reduction model, etc., associate the energy situation, and generate energy supply optimization suggestions;

[0099] Step 4: Link the park equipment management system and enterprise management system to issue energy supply optimization instructions.

[0100] Then, according to each step, the corresponding tools are scheduled or generated to complete the corresponding tasks. Finally, the output results are generated: energy policy analysis report, energy consumption analysis report, energy supply optimization report, energy supply adjustment (execution / non-execution / demonstration) and other results.

[0101] Through the above-mentioned fine-tuning learning and the introduction of experts to assist in optimization, the large language model can more accurately decompose tasks, making the final output results more in line with user needs.

[0102] Preferably, the step S5 also includes capability judgment and tool production. If it is judged that the tools dispatched by the language big model from the tool library and the functional models dispatched from the functional model library are poorly matched with the strategy solutions corresponding to the input tasks, tool generation is performed, and the language big model produces new tools with higher matching degree by itself, and then step S4 is re-executed to execute multiple subtasks, thereby further realizing the intelligence of the language big model and improving its versatility.

[0103] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the claims of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A multi-agent collaborative system suitable for industrial digitalization, comprising: A large model system for implementing central management functions; A tool library to provide basic functions; A functional model library for providing analytical operations; According to the input task, the large model system obtains and understands the user request corresponding to the task, decomposes the task into multiple goals and several subtasks corresponding to each goal, dispatches the tool corresponding to each subtask from the tool library, dispatches the functional model corresponding to each subtask from the functional model library, completes the combined calculation or decision of data and models, and forms a strategy plan corresponding to the input task; It is characterized in that it also includes a domain database related to the domain involved in the task and a memory module for realizing long and short-term memory; The large model system calls relevant data in the domain database according to the input task, so that the large model system can further understand the domain knowledge related to the task. The memory module is used to record the previous strategy scheme to provide a reference basis for the large model system to understand the context of the user request and the planning of each subtask; It also includes industrial business systems and comprehensive analysis and control systems; The industrial business system is used to provide the large model system with real-time operation data of the industry; The comprehensive analysis and control system is used to provide the large model system with periodic analysis data of the industry; By acquiring real-time industry operation data and periodic analysis data, the large model system can be refined to specific industrial parks when breaking down goals and tasks; It also includes a manual feedback module; The manual feedback module is used to feedback the prompt word input to the user, and the large model system further decomposes the task into multiple goals and a number of subtasks corresponding to each goal according to the prompt word feedback from the user; The artificial feedback module comprises: A knowledge base of all text generation task elements needed to complete a task; A vector database of all the text generation task descriptions mentioned above; A NOSQL database storing a text generation task description and text generation knowledge based on the text generation task description; The specific process of the manual feedback module for feeding back the prompt word input to the user is as follows: S41: The large model system obtains and understands the user request corresponding to the task and forms an initial prompt; S42: vectorizing the initial Prompt text, and matching the text generation task description that best meets the task requirements in the vector database of text generation task descriptions; S43: according to the text generation description, query the text generation knowledge that completes this task in the NOSQL database of the text generation task knowledge base; S44: constructing a system-level prompt, wherein the system-level prompt specifically refers to determining whether the user has fully described the text generation task when the large model system receives knowledge about the text generation task; S45: Based on the text generation knowledge obtained in S43 and the system-level prompt in S44, it is determined whether it is necessary to provide the user with a prompt word input; S46: if necessary, iterate the text of the prompt words fed back by the user to generate a task description; S47: Repeat S45 and S46 until the large model confirms that the user has completely described his text generation task, and then generates text based on the final text generation task description.

2. The multi-agent collaborative system suitable for industrial digitalization according to claim 1 is characterized in that: It also includes a capability judgment module, which is used to judge the matching degree between the tools scheduled from the tool library and the functional models scheduled from the functional model library by the large model system and the strategy solutions corresponding to the input tasks.

3. The multi-agent collaborative system suitable for industrial digitalization according to claim 2 is characterized in that: It also includes a tool production module. When the large model system determines through the capability judgment module that the selected tool cannot achieve the expected matching degree, the tool production module is called to produce a tool with a higher matching degree.

4. A strategy generation method for a multi-agent collaborative system suitable for industrial digitalization according to any one of claims 1 to 3, characterized in that: The following steps are involved: S1: Task planning: Use the language model to decompose the input task into multiple sub-goals; S2: Task parsing: Understand the user requests for multiple sub-goals based on the domain database related to the domain involved in the task and the strategy solutions previously recorded in the memory module, thereby generating multiple corresponding sub-tasks; S3: Tool and model selection: For each subtask, select the tool and functional model to implement the subtask; S4: Task execution: According to the selected tools and functional models, the corresponding subtasks are executed to achieve their respective sub-goals; S5: Strategy generation: Integrate multiple sub-goals to form an overall strategy solution.

5. The strategy generation method according to claim 4, characterized in that: Step S1 also includes manual feedback updating. If, when the language model receives knowledge about the text generation task, it is determined that the user has not fully described the task of his text, it is necessary to provide the user with prompt word input until the text entered by the user can fully describe the generated task and decompose the target into multiple sub-targets.

6. The strategy generation method according to claim 5, characterized in that: Step S2 also includes fine-tuning learning and expert optimization. Fine-tuning learning is performed by obtaining real-time operational data of the park industry and periodic data of the industry, so that the large language model can be refined to specific industrial parks when breaking down goals and tasks, and experts are introduced to assist in optimization.

7. The strategy generation method according to claim 6, characterized in that: The step S5 also includes capability judgment and tool production. If it is judged that the tools scheduled by the language big model from the tool library and the functional models scheduled from the functional model library are poorly matched with the strategy solutions corresponding to the input tasks, tool generation is performed, and the language big model produces new tools with higher matching degree by itself, and then step S4 is re-executed to execute multiple subtasks.

Images