LLM-based multi-agent collaborative decision-making system and decision-making method thereof

Abstract

According to the LLM-based multi-agent collaborative decision-making system and the decision-making method thereof, each agent of the decision-making system is configured with a schedulable path, and the agents are scheduled according to the schedulable paths without root routers. The decision-making method comprises the following steps of: 1, configuring a human prompt word, a schedulable path and a decision-making mechanism of each agent, and binding functions of the agents; 2, in the interaction process, when the question and answer service receives the user input content, the personal setting, the decision-making mechanism, the tool function description, the interaction context and the user input content are transmitted to the LLM together, so that the LLM becomes a central nervous system of an intelligent agent; and step 3, making one decision of scheduling, executing, replying / questioning by the intelligent agent. According to the invention, each agent can autonomously decide and dispatch without a root router, and the agent can be ensured to complete a specific task according to an artificially formulated execution process.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a multi-agent collaborative decision-making system based on LLM and a decision-making method thereof. Background Art

[0002] Large Language Model (LLM), also known as Large Language Model, is an artificial intelligence model designed to understand and generate human language. They are trained on large amounts of text data and can perform a wide range of tasks, including text summarization, translation, sentiment analysis, and more. LLMs are characterized by their large size, which helps them learn complex patterns in language data.

[0003] The rise of large language models has led to a huge change in the field of intelligent agent applications. Large language models have very good performance in logical reasoning, tool application, strategy planning, and instruction following. This has made people begin to realize that using large language models as the core cognitive system of intelligent agent applications can greatly improve the decision-making, scheduling, and planning capabilities of intelligent agents. In practical applications, intelligent agents often need to combine tools or external interfaces, or even multiple intelligent agents to assist each other in order to solve user tasks. How to fully utilize the autonomous decision-making and scheduling capabilities of intelligent agents, ensure that intelligent agents can complete specific tasks according to the execution process formulated by humans, and improve the accuracy and reliability of decision-making and scheduling has become a hot research direction at present.

[0004] Therefore, it is necessary to provide a multi-agent collaborative decision-making system and its decision-making method based on LLM to overcome the shortcomings of the existing technology. Summary of the invention

[0005] The purpose of the present invention is to avoid the shortcomings of the prior art and provide a multi-agent collaborative decision-making system and a decision-making method based on LLM, in which each agent can make decisions and schedule independently without going through a root router, and can ensure that the agent can complete specific tasks according to a manually formulated execution process.

[0006] The purpose of the present invention is achieved through the following technical measures.

[0007] A multi-agent collaborative decision-making system based on LLM is provided, which is provided with a root router and multiple agents, each of which is configured with a schedulable path. When the agent does not have the corresponding processing function itself, the schedulable path agent makes an autonomous decision to schedule to other agents with corresponding functional processing capabilities on the corresponding path; the scheduling between agents according to the schedulable path does not go through the root router.

[0008] Preferably, in the above-mentioned LLM-based multi-agent collaborative decision-making system, the schedulable path configured by an agent is the association relationship between the functions that the agent does not have and other agents that have corresponding functions.

[0009] Preferably, in the above-mentioned LLM-based multi-agent collaborative decision-making system, each agent is configured with a decision-making mechanism, which defines and manages the process of the agent processing tasks and guides and constrains the agent's decision-making.

[0010] Preferably, in the above-mentioned LLM-based multi-agent collaborative decision-making system, each agent is configured with functional information it possesses.

[0011] Preferably, in the above-mentioned LLM-based multi-agent collaborative decision-making system, the current agent autonomously executes decisions based on its own decision-making mechanism, the functions it possesses and its judgment of user needs.

[0012] Preferably, in the above-mentioned LLM-based multi-agent collaborative decision-making system, the agent uses LLM as the central nervous system for reasoning planning. Large Language Model (LLM), also known as Large Language Model, is an artificial intelligence model designed to understand and generate human language. They are trained on large amounts of text data and can perform a wide range of tasks, including text summarization, translation, sentiment analysis, and so on. LLMs are characterized by their large scale, which helps them learn complex patterns in language data.

[0013] Another object of the present invention is to provide a multi-agent collaborative decision-making method performed by the above-mentioned LLM-based multi-agent collaborative decision-making system, comprising the following steps:

[0014] Step 1. Configure or pre-configure the personality prompts, schedulable paths, decision-making mechanisms, and functions that the bound agents can have for each agent;

[0015] Step 2. During the interaction process, when the question-answering service receives the user input, it passes its own personality, decision-making mechanism, tool function description, interaction context and user input content to the LLM, making it the central nervous system of the intelligent agent;

[0016] Step 3. The agent makes one of the following decisions: schedule, execute, reply / ask.

[0017] Preferably, in the above-mentioned multi-agent collaborative decision-making method, step 1 configures different personality prompt word descriptions for each agent, so that the agent is aware of its own capabilities and the problems it can solve.

[0018] Preferably, in the above-mentioned multi-agent collaborative decision-making method, in step 3, when the agent believes that the functions it possesses are unable to solve the current task, the agent makes a scheduling decision and hands over the task and the executed results to other agents for resolution according to the schedulable path.

[0019] Preferably, in the above multi-agent collaborative decision-making method, in step 3, the execution decision is specifically: based on the current task, the agent uses the tools it has to process and obtain the execution result; and the execution result is added to the interactive context information, and the current agent re-decides until the reply ends the cycle;

[0020] The reply / question decision is as follows: if the agent believes that it can give the user a clear answer based on the known information, it will reply directly; or if it believes that the user needs to provide more detailed information, it will ask the user a question and interact with the user.

[0021] The multi-agent collaborative decision-making system and decision-making method based on LLM of the present invention simplify the scheduling process between agents, so that the user demand task does not need to go through the root route during execution, and can autonomously execute the decision based on the current agent according to its own decision-making mechanism, functions and judgment of needs. Secondly, the detailed process of agent processing tasks is defined and managed through artificial decision-making mechanisms to achieve the purpose of guiding and constraining the decision-making of agents. The agent learns about its functions and schedulable paths through configuration items. The agent makes decisions (scheduling / execution / reply) based on the user demand task, and during this period, the decision record and the execution result of the decision will be added to the context and passed on until LLM considers that the user's question can be answered.

[0022] The multi-agent collaborative decision-making system and decision-making method based on LLM of the present invention have the following highlights:

[0023] Highlight 1: Each agent is a router itself and can make its own decisions on whether to schedule, execute or reply, rather than relying on the root router for scheduling and distribution.

[0024] Highlight 2: Each agent is guided and constrained by the decision-making mechanism, which serves as the standard strategy for its task execution scheduling.

[0025] Highlight 3: The agent’s personality, schedulable paths, decision-making mechanism, and tools / external interfaces are all manually designed. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention is further described with reference to the accompanying drawings, but the contents in the accompanying drawings do not constitute any limitation to the present invention.

[0027] Figure 1 It is a schematic diagram of a multi-agent collaborative decision-making system based on LLM of the present invention.

[0028] Figure 2 It is a schematic diagram of the basic framework of the intelligent agent. DETAILED DESCRIPTION

[0029] The present invention will be further described with reference to the following examples.

[0030] Example 1.

[0031] A multi-agent collaborative decision-making system based on LLM, such as Figure 1 As shown, a root router and multiple intelligent agents are set up, each intelligent agent is configured with a schedulable path, and through the schedulable path intelligent agent can autonomously decide to schedule to other intelligent agents with corresponding functional processing capabilities on the corresponding path when it does not have the corresponding processing function itself; the scheduling between intelligent agents according to the schedulable path does not go through the root router.

[0032] In this LLM-based multi-agent collaborative decision-making system, the schedulable path configured by an agent is the association between the functions that the agent does not have and other agents that have corresponding functions. The current agent autonomously executes decisions based on its own decision-making mechanism, the functions it has, and the judgment of user needs. By setting the schedulable path, the scheduling process between agents is simplified, so that user demand tasks do not need to go through the root route during execution, and can autonomously execute decisions based on the current agent's own decision-making mechanism, the functions it has, and the judgment of needs.

[0033] Each agent is configured with a decision-making mechanism, which defines and manages the process of the agent processing tasks and guides and constrains the agent's decision-making. Each agent is configured with functional information. The detailed process of the agent processing tasks is defined and managed by artificially formulating a decision-making mechanism to achieve the purpose of guiding and constraining the agent's decision-making. The agent learns its functions and schedulable paths through configuration items. The agent makes decisions (scheduling / execution / reply) based on the user's required tasks.

[0034] In this LLM-based multi-agent collaborative decision-making system, the agents use LLM as the central nervous system for reasoning and planning. Large Language Model (LLM), also known as Large Language Model, is an artificial intelligence model designed to understand and generate human language. They are trained on large amounts of text data and can perform a wide range of tasks, including text summarization, translation, sentiment analysis, and more. LLMs are characterized by their large scale, which helps them learn complex patterns in language data.

[0035] like Figure 2As shown, an intelligent agent = LLM + memory + tools + routing + decision-making + planning, where LLM plays the role of the central nervous system of the agent and provides the ability of reasoning and planning.

[0036] Memory, including short-term memory and long-term memory.

[0037] Short-term memory: Agent personality, decision-making mechanism, historical interaction context, and tool descriptions all belong to short-term memory.

[0038] Long-term memory: Provides the agent with the ability to retain and recall unlimited information for a long time, usually by utilizing external vector storage and fast retrieval.

[0039] Tools: The biggest difference between Agent and LLM is whether it has tools. As an important part of Agent's interaction with the outside world, Agent can use external tools to expand model capabilities and complete task requirements by combining the execution of different tools.

[0040] Routing: The paths that each agent can use to jump to other agents are artificially defined.

[0041] Decision-making: Decisions made by the agent based on short-term memory, including scheduling, execution, and response.

[0042] Scheduling: The agent believes that its own functions are not capable of solving the current task and needs to hand over the task and the executed results to other agents for resolution.

[0043] Execution: Based on the current task, the agent uses its own tools to process and obtain the execution result. The execution result is added to the interaction context information, and the LLM makes a new decision until the reply ends the loop.

[0044] Reply: When the agent believes that it can give the user a clear answer based on the known information, it will reply directly.

[0045] Planning: Planning includes task decomposition and reflection and refinement. There are two main implementation modes: CoT and ReAct.

[0046] Sub-goals and decomposition: Complex tasks can be handled efficiently by breaking them down into simple sub-goals and developing execution steps for each sub-goal.

[0047] Self-reflection: Agents can reflect on past actions, learn from mistakes, and improve future steps, thereby improving the quality of the final result.

[0048] Chain of Thought (CoT): The performance of a large model can be significantly improved by having it decompose a complex problem into sub-problems step by step and solve them one by one. The intermediate steps of this series of reasoning are called Chain of Thought and have become a standard prompting technique for improving the performance of models in complex tasks. The model is asked to "think step by step" to break down difficult tasks into smaller and simpler steps.

[0049] ReAct: It still uses the strategy of divide-and-conquer for complex tasks, but unlike the mind chain, ReAct is executed through multiple iterations, and the results of each execution will serve as context information to prompt LLM to dynamically adjust its decision.

[0050] The multi-agent collaborative decision-making method of the multi-agent collaborative decision-making system based on LLM has the following steps:

[0051] Step 1. Configure or pre-configure the personality prompts, schedulable paths, decision-making mechanisms, and functions that each agent can have. Step 1 configures different personality prompts for each agent so that the agent knows its own capabilities and the problems it can solve.

[0052] Scheduling paths: By configuring schedulable paths, the paths that each agent can jump to other agent nodes are limited. This reduces the tokens consumed for each decision and further improves the accuracy and efficiency of decisions. The root routing node is standard for each agent. When faced with complex tasks, it can serve as the decision-making brain and information transfer station for multiple agents, planning the collaborative execution of multiple agents to solve problems and improve the overall planning capabilities of the system.

[0053] Decision-making mechanism: Standard specifications for decision-making are established for each agent, which can guide and constrain the execution process of the agent when solving specific problems, ensure that tasks are executed according to established processes, improve overall execution efficiency, and avoid executing complex decision-making processes due to the agent's own misunderstanding of the task.

[0054] Tools / external interfaces: tools or external interfaces that need to be implemented by users. Common tools include calculators, and external interfaces include weather / search engines / route planning queries, etc. Depending on the functions of each agent, the tasks it can solve are also different.

[0055] Step 2. During the interaction process, when the question-answering service receives user input, it passes its own personality, decision-making mechanism, tool function description, interaction context and user input content to the LLM, making it the central nervous system of the intelligent agent.

[0056] Step 3. The agent makes one of the following decisions: schedule, execute, reply / ask.

[0057] When the agent believes that the functions it possesses are unable to solve the current task, the agent makes a scheduling decision and hands over the task and the executed results to other agents for resolution according to the schedulable path.

[0058] The specific execution decision is: based on the current task, the agent uses the tools it has to process and obtain the execution result; the execution result is added to the interaction context information, and the current agent makes a new decision until the reply ends the cycle;

[0059] The reply / question decision is as follows: if the agent believes that it can give the user a clear answer based on the known information, it will reply directly; or if it believes that the user needs to provide more detailed information, it will ask the user a question and interact with the user.

[0060] The multi-agent collaborative decision-making system and decision-making method based on LLM of the present invention simplify the scheduling process between agents, so that the user demand task does not need to go through the root route during execution, and can autonomously execute the decision based on the current agent according to its own decision-making mechanism, functions and judgment of needs. Secondly, the detailed process of agent processing tasks is defined and managed through artificial decision-making mechanisms to achieve the purpose of guiding and constraining the decision-making of agents. The agent learns about its functions and schedulable paths through configuration items. The agent makes decisions (scheduling / execution / reply) based on the user demand task, and during this period, the decision record and the execution result of the decision will be added to the context and passed on until LLM considers that the user's question can be answered.

[0061] In the LLM-based multi-agent collaborative decision-making system and its decision-making method, each agent is a route itself and can decide whether to schedule, execute or reply, rather than relying on the root route scheduling distribution. Each agent is guided and constrained by the decision-making mechanism, which serves as the standard strategy for its task execution scheduling. The agent settings, schedulable paths, decision-making mechanisms, and tools / external interfaces are all manually formulated.

[0062] Example 2.

[0063] The present invention illustrates a multi-agent collaborative decision-making system based on LLM and its decision-making method through a specific example.

[0064] In this embodiment, a specific example of multi-agents assisting users in completing itinerary arrangements is given, wherein the multi-agents include a route planning agent, a food recommendation agent, and an accommodation recommendation agent.

[0065] Step 1: Users need to configure each agent with its own personality prompts, schedulable paths, decision-making mechanisms, and bound functions (tools / external interfaces).

[0066] Route planning agent-personnel settings are as follows:

[0067] “You are an intelligent transportation assistant, and you need to provide detailed travel plans based on the user’s travel needs.

[0068] If there are multiple travel options, highlight the different train or flight schedules and arrival times.

[0069] Before starting each strategy, read all messages from users and the entire strategy steps.

[0070] Strictly follow the following policies.

[0071] If you are unsure about the next step in the strategy traversal, ask the client for more information.

[0072] Important: Never share contextual details with users.

[0073] Important: You must always complete all steps in a strategy before continuing.

[0074] Important: Start with step one of your strategy now! "

[0075] Route planning agent-decision-making mechanism is as follows:

[0076] “1. According to the user’s question, determine the user’s departure place, destination and specific travel date (%Y-%m-%d), and plan the route.

[0077] a). If the user's chosen mode of travel is airplane, call the search_flight function;

[0078] b). If the user selects the train as the mode of travel, call the search_train function;

[0079] 2. Show users the latest travel options;

[0080] 3. If the user needs hotel accommodation recommendations, call the transfer_to_accommodation function;

[0081] 4. If the user needs food recommendations, call the transfer_to_delicacy function. ”

[0082] Route Planning Agent - Scheduling Paths:

[0083] The schedulable path is passed to LLM in the form of a function, and the schedulable paths possessed by different agents are not consistent.

[0084] "transfer_to_triage#Transfer the user to the root routing agent (scheduling)transfer_to_travel_manage#Transfer the user to the travel management agent (scheduling)transfer_to_delicacy#Jump to the food recommendation agent (scheduling)transfer_to_accommodation#Jump to the food recommendation agent (scheduling)".

[0085] Route Planning Agent - Tools / External Interfaces:

[0086] The configuration method is consistent with the schedulable path, and is passed to the LLM in the form of a function.

[0087] Users need to implement specific functions, encapsulate them into executable business functions, and finally convert them into tool function descriptions that LLM can understand, so that LLM can extract the required function parameters from the interactive context based on the function description.

[0088] search_flight#Search for flight travel options (execution)

[0089] search_train#Search for train travel options (execution)".

[0090] Executable business functions (the following only lists the function of dispatching to the itinerary management agent and the encapsulation function of querying the train route. Since the tool function is not the focus of this embodiment, the specific implementation details of the function are not further described here.)

[0091] Dispatching to the trip management function:

[0092] def transfer_to_travel_manage():return travel_manage_agent.

[0093] Query train route function:

[0094] "def search_train(origin,destination,departure_date):

[0095] """

[0096] Check train travel options

[0097] parameter:

[0098] origin(str): Detailed origin.

[0099] destination(str): Detailed destination.

[0100] departure_date(str): departure date.

[0101] return:

[0102] dict:Query train travel options

[0103] """

[0104] Import requests

[0105] response=requests.get(f"http: / / {ip}:8082 / train_routes_planning",params={"origin":origin,"destination":destination,"departure_date":departure_date})

[0106] return response.json()['routes_result']".

[0107] Tool function descriptions that LLM can understand (convert the function into json descriptions, which are then converted into instructions for using the function, including function parameters and returns, to help LLM understand the functions and calling methods of different tool functions).

[0108] {

[0109] "type":"function",

[0110] "function":{

[0111] "name":"search_train",

[0112] "description":"\nQuery train travel plans\n\nParameters:\n origin(str): detailed origin.\n destination(str): detailed destination.\n departure_date(str): departure date.\n\nReturn:\n dict: Query train travel plans\n",

[0113] "parameters":{

[0114]

[0115] Personal Assistant Agent:

[0116] Character: "You are a professional personal secretary, responsible for unified management of the user's itinerary. This includes recommending travel plans, planning itineraries, adding new itineraries, querying all the user's itineraries, querying the specific itinerary arrangements for a certain day, modifying the itinerary date, travel content, departure place, destination, travel plan, food reservations, hotel accommodation reservations, etc.

[0117] Before starting each strategy, read all messages from users and the entire strategy steps.

[0118] When answering each question, you need to strictly follow the following strategy.

[0119] Only when you get to the point where you can call case_resolved, and have confirmed that the customer has no further questions, can the strategy be considered complete.

[0120] If you are unsure about the next step in the strategy traversal, ask the client for more information. Important: Never share contextual details with the user.

[0121] Important: You must always complete all steps in a strategy before continuing.

[0122] Important: Please do not share the details of the task transfer in your answer.

[0123] Note: If the user's name or travel dates are unknown, ask the user.

[0124] NOTE: Always call the 'transfer_to_triage' function if the user's requirements are no longer relevant to the selected strategy. IMPORTANT: Start immediately from the first step of the strategy! "

[0125] Decision-making mechanism:

[0126] “1. If you are not sure about the exact departure location or travel date, ask the user.

[0127] 2. If the user wants to query all of his trips, call the search_travel function.

[0128] 3. If the user wants to query the specific itinerary for a certain day, call the search_travel_details function.

[0129] 4. If the user needs to make an itinerary or travel plan, follow the following process one by one: 4-1). If the travel date is uncertain, ask for the travel date and call the modify_travel_date function; proceed to the next step;

[0130] 4-2). If you are not sure about the itinerary, ask about itinerary content and call modify_travel_aim function; go to the next step;

[0131] 4-3). If the departure place of the trip is not sure, ask the departure place of the trip and call the modify_travel_origin function; then go to the next step;

[0132] 4-4). If the destination is not certain, ask for the destination and call the modify_travel_destination function; go to the next step;

[0133] 4-5). According to the departure and destination, call the transfer_to_traffic function; go to the next step; 4-6). Ask the user if he needs food or accommodation suggestions;

[0134] 4-7). If the user needs food recommendations, call the transfer_to_delicacy function; 4-8). If the user needs hotel accommodation recommendations, call the transfer_to_accommodation function;

[0135] 5. If the user needs to change the itinerary, choose to change it according to the user's needs:

[0136] a). If the user changes the travel date, directly execute the transfer_to_traffic function to re-plan the travel plan.

[0137] 1a). Call modify_travel_date function to change the travel date. Go to the next step;

[0138] 2a). Call transfer_to_accommodation function; re-plan the accommodation plan; proceed to the next step;

[0139] 3a). Call modify_travel_accommodation function to change accommodation.

[0140] b). If the user changes the departure point, directly execute the transfer_to_traffic function and re-plan the travel plan.

[0141] 1b). Call modify_travel_origin function to change the departure place.

[0142] c). If the user changes the destination, follow the steps below:

[0143] 1c). Execute the transfer_to_traffic function directly to re-plan the travel plan. Go to the next step; 2c). Call the modify_travel_destination function to change the destination. Go to the next step; 3c). Call the modify_travel_delicacy function to change the delicacy. Go to the next step;

[0144] 4c). Call modify_travel_accommodation function to change accommodation.

[0145] d). If the user changes the travel plan, directly execute the transfer_to_traffic function to re-plan the travel plan.

[0146] e). If the user changes the cuisine, call the modify_travel_delicacy function;

[0147] f). If the user changes accommodation, call the modify_travel_accommodation function;".

[0148] Step 2: During the interaction with the user, after receiving the user input, the question-answering service will pass its own personality, decision-making mechanism, tool function description, interaction context and user input content to the LLM, making it an intelligent agent.

[0149] Step 3: Based on the above information, the agent makes one of the following decisions: scheduling, execution, or reply / question:

[0150] Scheduling: The agent believes that its own functions are not capable of solving the current task and needs to hand over the task and the executed results to other agents for resolution.

[0151] Execution: Based on the current task, the agent uses the tools it has to process and obtain the execution result. The execution result is added to the interaction context information, and the current agent makes a new decision until the reply ends the loop.

[0152] Reply / ask questions: If the agent thinks it can give the user a clear answer based on the known information, it will reply directly. Or if it thinks the user needs to provide more detailed information, it will ask questions to interact with the user.

[0153] For example, the user inputs: I am going to Beijing tomorrow, please help me plan my itinerary.

[0154] After the multi-agent receives the user input, the root routing agent enters and starts execution. The root routing agent believes that it does not have the ability to solve the task based on its own personality, decision-making mechanism, tool function description and user input, so it calls the [Scheduling] function in the tool function to jump to the itinerary planning agent.

[0155] Based on its own personality, decision-making mechanism, tool function description and context, the itinerary planning agent believes that the current task needs to be solved by calling the route planning tool, but the user needs to provide the departure point. At this time, it will make a [Ask] decision, ask the user for the departure point, and update the question to the interaction context.

[0156] Itinerary Planning Agent: OK, you need to provide the departure location so that I can plan a suitable travel route for you.

[0157] User: I am leaving from Guangzhou.

[0158] Get the user's reply, update it to the interaction context, and input it to the itinerary planning agent again. The itinerary planning agent believes that it is necessary to [execute] the search_train tool function it has to plan the travel route. According to the function description of search_train, extract the starting point and destination parameters in the context.

[0159] The extracted parameters are: {"origin":"Beijing","destination":"Shanghai"}.

[0160] And execute the search_train function to get the following function return.

[0161] {

[0162] "routes_result":{

[0163] "airplane":[

[0164] "[Bus] Take Metro Line 2 (toward Jiahe Wanggang) from Gongyuanqian Station (G Animation Star City Exit) and get off at Jiahe Wanggang Station after 11 stops\n[Bus] Take Metro Line 3 (toward Airport North (Terminal 2)) from Jiahe Wanggang Station and get off at Airport South (Terminal 1) Station after 4 stops\n[Airplane] Take KY9187 from Guangzhou Baiyun Airport to Beijing Capital Airport\n\t->Fare: 1480￥\n\t->Departure time: 2024-11-01

[0165] 07:30:00\n\t-> Arrival time: 2024-11-0110:20:00\n[Bus] Take the Capital Airport Line (towards Beixinqiao) from Terminal 2 Station (subway entrance-Parking Lot 2) and get off at Dongzhimen Station after 2 stops\n[Bus] Take the Inner Ring Line of Subway Line 2 (towards Dongsishitiao) from Dongzhimen Station and get off at Jianguomen Station after 3 stops\n[Bus] Take the Batong Line of Subway Line 1 (towards Gucheng) from Jianguomen Station and get off at Wangfujing Station (Southwest Exit of C4) after 2 stops",

[0166] "[Bus] Take Metro Line 2 (toward Jiahe Wanggang) from Gongyuanqian Station (G Animation Star City Exit) and get off at Jiahe Wanggang Station after 11 stops\n[Bus] Take Metro Line 3 (toward Airport North (Terminal 2)) from Jiahe Wanggang Station and get off at Airport South (Terminal 1) Station after 4 stops\n[Airplane] Take TV6473 from Guangzhou Baiyun Airport to Beijing Capital Airport\n\t->Fare: 1480￥\n\t->Departure time: 2024-11-01

[0167] 07:30:00\n\t-> Arrival time: 2024-11-0110:20:00\n[Bus] Take the Capital Airport Line (towards Beixinqiao) from Terminal 2 Station (subway entrance-Parking Lot 2) and get off at Dongzhimen Station after 2 stops\n[Bus] Take the Inner Ring Line of Subway Line 2 (towards Dongsishitiao) from Dongzhimen Station and get off at Jianguomen Station after 3 stops\n[Bus] Take the Batong Line of Subway Line 1 (towards Gucheng) from Jianguomen Station and get off at Wangfujing Station (Southwest Exit of C4) after 2 stops",

[0168] "[Bus] Take Metro Line 2 (toward Jiahe Wanggang) from Gongyuanqian Station (G Animation Star City Exit) and get off at Jiahe Wanggang Station after 11 stops\n[Bus] Take Metro Line 3 (toward Airport North (Terminal 2)) from Jiahe Wanggang Station and get off at Airport South (Terminal 1) Station after 4 stops\n[Airplane] Take CA3559 from Guangzhou Baiyun Airport to Beijing Capital Airport\n\t->Fare: 1480￥\n\t->Departure time: 2024-11-01

[0169] 07:30:00\n\t-> Arrival time: 2024-11-0110:20:00\n[Bus] Take the Capital Airport Line (towards Beixinqiao) from Terminal 2 Station (subway entrance-Parking Lot 2) and get off at Dongzhimen Station after 2 stops\n[Bus] Take the Inner Ring Line of Subway Line 2 (towards Dongsishitiao) from Dongzhimen Station and get off at Jianguomen Station after 3 stops\n[Bus] Take the Batong Line of Subway Line 1 (towards Gucheng) from Jianguomen Station and get off at Wangfujing Station (Southwest Exit of C4) after 2 stops",

[0170] "[Bus] Take Metro Line 2 (toward Jiahe Wanggang) from Gongyuanqian Station (G Animation Star City Exit) and get off at Jiahe Wanggang Station after 11 stops\n[Bus] Take Metro Line 3 (toward Airport North (Terminal 2)) from Jiahe Wanggang Station and get off at Airport South (Terminal 1) Station after 4 stops\n[Airplane] Take ZH9147 from Guangzhou Baiyun Airport to Beijing Capital Airport\n\t->Fare: 935￥\n\t->Departure time: 2024-11-01

[0171] 07:30:00\n\t-> Arrival time: 2024-11-0110:20:00\n[Bus] Take the Capital Airport Line (towards Beixinqiao) from Terminal 2 Station (subway entrance-Parking Lot 2) and get off at Dongzhimen Station after 2 stops\n[Bus] Take the Inner Ring Line of Subway Line 2 (towards Dongsishitiao) from Dongzhimen Station and get off at Jianguomen Station after 3 stops\n[Bus] Take the Batong Line of Subway Line 1 (towards Gucheng) from Jianguomen Station and get off at Wangfujing Station (Southwest Exit of C4) after 2 stops",

[0172] "[Bus] Take Metro Line 2 (toward Jiahe Wanggang) from Gongyuanqian Station (G Animation Star City Exit) and get off at Jiahe Wanggang Station after 11 stops\n[Bus] Take Metro Line 3 (toward Airport North (Terminal 2)) from Jiahe Wanggang Station and get off at Airport South (Terminal 1) Station after 4 stops\n[Airplane] Take SC9197 from Guangzhou Baiyun Airport to Beijing Capital Airport\n\t->Fare: 1480￥\n\t->Departure time: 2024-11-01

[0173] 07:30:00\n\t-> Arrival time: 2024-11-0110:20:00\n[Bus] Take the Capital Airport Line (towards Beixinqiao) from Terminal 2 Station (subway exit-Parking Lot 2) and get off at Dongzhimen Station after 2 stops\n[Bus] Take the Inner Ring Line of Subway Line 2 (towards Dongsishitiao) from Dongzhimen Station and get off at Jianguomen Station after 3 stops\n[Bus] Take the Batong Line of Subway Line 1 (towards Gucheng) from Jianguomen Station and get off at Wangfujing Station (Southwest Exit of C4) after 2 stops"

[0174] ],

[0175] "train":[

[0176] "[Bus] Take Metro Line 2 (towards Guangzhou South Station) from Gongyuanqian Station (Exit G) and get off at Guangzhou South Station (Exit F) after 12 stops\n[Train] Take G82 from Guangzhou South Station to Beijing West Station

[0177] \n\t->Ticket price: 1033￥\n\t->Departure time: 2024-11-0108:38:00\n\t->Arrival time: 2024-11-0116:27:00\n[Public transportation] Take Subway Line 9 (towards the National Library) from Beijing West Railway Station (Southwest Exit 4) and get off at the Military Museum Station after 1 stop\n[Public transportation] Take Subway Line 1 Batong Line (towards Universal Studios) from the Military Museum Station and get off at Wangfujing Station (Southwest Exit C4) after 7 stops",

[0178] "[Bus] Take Metro Line 2 (towards Guangzhou South Station) from Gongyuanqian Station (Exit G) and get off at Guangzhou South Station (Exit F) after 12 stops\n[Train] Take G336 from Guangzhou South Station to Beijing West Station

[0179] \n\t->Fare: 862￥\n\t->Departure time: 2024-11-0107:22:00\n\t->Arrival time: 2024-11-0117:58:00\n[Public transportation] Take Subway Line 9 (towards the National Library) from Beijing West Railway Station (Southwest Exit 4) and get off at the Military Museum Station after 1 stop\n[Public transportation] Take Subway Line 1 Batong Line (towards Universal Studios) from the Military Museum Station and get off at Wangfujing Station (Southwest Exit C4) after 7 stops",

[0180] "[Bus] Take Metro Line 2 (towards Guangzhou South Station) from Gongyuanqian Station (Exit G) and get off at Guangzhou South Station (Exit F) after 12 stops\n[Train] Take G1580 from Guangzhou South Station to Beijing West Station

[0181] \n\t->Fare: 775￥\n\t->Departure time: 2024-11-0108:10:00\n\t->Arrival time: 2024-11-0119:13:00\n[Public transportation] Take Subway Line 9 (towards the National Library) from Beijing West Railway Station (Southwest Exit 4) and get off at the Military Museum Station after 1 stop\n[Public transportation] Take Subway Line 1 Batong Line (towards Universal Studios) from the Military Museum Station and get off at Wangfujing Station (Southwest Exit C4) after 7 stops",

[0182] "[Bus] Take Metro Line 2 (towards Guangzhou South Station) from Gongyuanqian Station (Exit G) and get off at Guangzhou South Station (Exit F) after 12 stops\n[Train] Take G338 from Guangzhou South Station to Beijing West Station

[0183] \n\t->Fare: 862￥\n\t->Departure time: 2024-11-0111:15:00\n\t->Arrival time: 2024-11-0121:20:00\n[Public transportation] Take Subway Line 9 (towards the National Library) from Beijing West Railway Station (Southwest Exit 4) and get off at the Military Museum Station after 1 stop\n[Public transportation] Take Subway Line 1 Batong Line (towards Universal Studios) from the Military Museum Station and get off at Wangfujing Station (Southwest Exit C4) after 7 stops",

[0184] "[Bus] Take Metro Line 2 (towards Guangzhou South Station) from Gongyuanqian Station (G Animation Star City Exit) and get off at Guangzhou South Station (Exit F) after 12 stops\n[Train] Take G78 from Guangzhou South Station to Beijing Station\n\t->Fare: 1036￥\n\t->Departure time: 2024-11-0114:16:00\n\t->Arrival time: 2024-11-0122:12:00\n[Bus] Take bus No. 622 (towards Chengheyuan Community) from Beijing Station East Station and get off at Zhengyi Road Station after 3 stops"

[0185] ],

[0186] "Bus":[],

[0187] "Beijing":"City: Beijing, Region: Beijing, Date: 2024-12-01, Highest temperature: Unknown, Lowest temperature: Unknown, Daytime weather: Unknown, Nighttime weather: Unknown"

[0188] }

[0189] }

[0190] After the trip planning agent executes the tool and updates the context, it believes that it can answer the user's question based on the current information. It re-decides and selects [Reply] to reply to the user. The reply content is as follows:

[0191] "According to the query results, you have the following flight travel options from Guangzhou to Beijing

[0192] 1. Flight: KY9187

[0193] Ticket price: 1480 yuan

[0194] Departure time: 2024-11-26 07:30

[0195] Arrival time:2024-11-2610:20

[0196] Departure Airport: Guangzhou Baiyun Airport

[0197] Arrival Airport: Beijing Capital Airport

[0198] 2. Flight: TV6473

[0199] Ticket price: 1480 yuan

[0200] Departure time: 2024-11-26 07:30

[0201] Arrival time:2024-11-2610:20

[0202] Departure Airport: Guangzhou Baiyun Airport

[0203] Arrival Airport: Beijing Capital Airport

[0204] 3. Flight: CA3559

[0205] Ticket price: 1480 yuan

[0206] Departure time: 2024-11-26 07:30

[0207] Arrival time:2024-11-2610:20

[0208] Departure Airport: Guangzhou Baiyun Airport

[0209] Arrival Airport: Beijing Capital Airport

[0210] 4. Flight: ZH9147

[0211] Ticket price: 935 yuan

[0212] Departure time: 2024-11-26 07:30

[0213] Arrival time:2024-11-2610:20

[0214] Departure Airport: Guangzhou Baiyun Airport

[0215] Arrival Airport: Beijing Capital Airport

[0216] 5. Flight: SC9197

[0217] Ticket price: 1480 yuan

[0218] Departure time: 2024-11-26 07:30

[0219] Arrival time:2024-11-2610:20

[0220] Departure Airport: Guangzhou Baiyun Airport

[0221] Arrival Airport: Beijing Capital Airport

[0222] Please select a flight that suits you and I will save your travel plan for you. "

[0223] The multi-agent collaborative decision-making system and decision-making method based on LLM of the present invention simplify the scheduling process between agents, so that the user demand task does not need to go through the root route during execution, and can autonomously execute the decision based on the current agent according to its own decision-making mechanism, functions and judgment of needs. Secondly, the detailed process of agent processing tasks is defined and managed through artificial decision-making mechanisms to achieve the purpose of guiding and constraining the decision-making of agents. The agent learns about its functions and schedulable paths through configuration items. The agent makes decisions (scheduling / execution / reply) based on the user demand task, and during this period, the decision record and the execution result of the decision will be added to the context and passed on until LLM considers that the user's question can be answered.

[0224] In the LLM-based multi-agent collaborative decision-making system and its decision-making method, each agent is a route itself and can decide whether to schedule, execute or reply, rather than relying on the root route scheduling distribution. Each agent is guided and constrained by the decision-making mechanism, which serves as the standard strategy for its task execution scheduling. The agent settings, schedulable paths, decision-making mechanisms, and tools / external interfaces are all manually formulated.

[0225] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. A multi-agent collaborative decision-making system based on LLM, provided with a root router and multiple agents, characterized in that: Each intelligent agent is configured with a schedulable path. Through the schedulable path, the intelligent agent can make an autonomous decision to schedule to other intelligent agents with corresponding functional processing capabilities on the corresponding path when it does not have the corresponding processing function itself; the scheduling between intelligent agents based on the schedulable path does not go through the root router.

2. The multi-agent collaborative decision-making system based on LLM according to claim 1 is characterized in that: The schedulable path of an agent configuration is the association between the functions that the agent does not have and other agents that have corresponding functions.

3. The multi-agent collaborative decision-making system based on LLM according to claim 2 is characterized in that: Each agent is configured with a decision-making mechanism, which defines and manages the process of the agent processing tasks and guides and constrains the agent's decision-making.

4. The multi-agent collaborative decision-making system based on LLM according to claim 3 is characterized in that: Each agent is configured with information about its capabilities.

5. The LLM-based multi-agent collaborative decision-making system according to claim 4 is characterized in that: The current intelligent agent makes decisions autonomously based on its own decision-making mechanism, the functions it possesses, and its judgment of user needs.

6. The LLM-based multi-agent collaborative decision-making system according to claim 5 is characterized by: The agent uses LLM as the central nervous system for reasoning and planning.

7. A multi-agent collaborative decision-making method performed by a multi-agent collaborative decision-making system based on LLM as claimed in any one of claims 1 to 6, characterized in that: The steps are as follows: Step 1. Configure or pre-configure the personality prompts, schedulable paths, decision-making mechanisms, and functions that the bound agents can have for each agent; Step 2. During the interaction process, when the question-answering service receives the user input, it passes its own personality, decision-making mechanism, tool function description, interaction context and user input content to the LLM, making it the central nervous system of the intelligent agent; Step 3. The agent makes one of the following decisions: schedule, execute, reply / ask.

8. The multi-agent collaborative decision-making method according to claim 7, characterized in that: Step 1 configures different personality prompt word descriptions for each agent so that the agent can understand its own capabilities and the problems it can solve.

9. The multi-agent collaborative decision-making method according to claim 8, characterized in that: In step 3, when the agent believes that its functions are unable to solve the current task, the agent makes a scheduling decision and hands over the task and the executed results to other agents for resolution according to the schedulable path.

10. The multi-agent collaborative decision-making method according to claim 8, characterized in that: In step 3, the execution decision is as follows: based on the current task, the agent uses the tools it has to process and obtain the execution result; the execution result is added to the interaction context information, and the current agent makes a new decision until the reply ends the loop; The reply / question decision is as follows: if the agent believes that it can give the user a clear answer based on the known information, it will reply directly; or if it believes that the user needs to provide more detailed information, it will ask the user a question and interact with the user.

Images