A collaborative method and apparatus for multi-agent systems under information asymmetry conditions

Abstract

This invention provides a collaborative method and apparatus for a multi-agent system under information asymmetry. The method includes: receiving an input task; generating unknown inference criteria and known inference criteria based on the input task and pre-acquired dialogue information between the target agent and at least one counterpart target agent; generating a question based on the unknown inference criteria and sending it to the counterpart target agent, and updating the corresponding unknown inference criteria to known inference criteria based on the response returned by the counterpart target agent; collecting the final updated known inference criteria of the target agent and the final updated known inference criteria of the counterpart target agent to perform consensus inference to obtain the final answer for responding to the task, thereby solving the information asymmetry problem, ensuring that each target agent can obtain all the information needed to complete the task, and achieving more efficient multi-agent collaboration and information processing capabilities.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a collaborative method and apparatus for multi-agent systems under conditions of information asymmetry. Background Technology

[0002] With the development of artificial intelligence technology, multi-agent systems (MAS) based on large language models (LLM) have made significant progress in solving complex tasks. These systems achieve collaborative work by communicating natural language between agents.

[0003] In existing technologies, when multi-agent systems are used to enhance cooperation among users, each agent can only access information from its own users. This leads to a situation where, when dealing with information asymmetry, agents cannot obtain the complete information required for the task, thus hindering accurate decision-making. This severely limits the efficiency of multi-agent systems.

[0004] For example, in a virtual team, different members' agents may each hold different parts of the project's information, which is crucial to the project's success. If information cannot be effectively exchanged between agents, the knowledge and skills of each member cannot be fully utilized, leading to inefficient collaboration.

[0005] In summary, existing technologies have significant shortcomings in handling information asymmetry problems, and a new approach is needed to improve the collaboration efficiency of multi-agent systems under information asymmetry conditions. Summary of the Invention

[0006] This invention provides a collaborative method and apparatus for multi-agent systems under information asymmetry, which solves the problem that in the prior art, agents in multi-agent systems cannot obtain complete information required for the task under information asymmetry, thus failing to make accurate decisions and severely limiting the efficiency of multi-agent systems.

[0007] This invention provides a collaborative method for a multi-agent system in situations of information asymmetry. The multi-agent system includes a relational network formed by multiple agents based on a large language model. For any target agent: the target agent and at least one peer target agent connected to it form a relational sub-network; each target agent pre-obtains the dialogue information of at least one peer target agent in the same relational sub-network. The method is used for the target agent and includes the following steps: The task of receiving input; Based on the input task and the pre-acquired dialogue information between the target agent and at least one opposing target agent, generate unknown reasoning basis and known reasoning basis; Based on the unknown reasoning basis, a question is generated and sent to the target agent. The unknown reasoning basis is updated to a known reasoning basis based on the response from the target agent. Collect the known reasoning basis for the final update completion of the target agent and the known reasoning basis for the final update completion of the opposing target agent, perform consensus reasoning, and obtain the final answer for responding to the task.

[0008] According to the present invention, a collaborative method for a multi-agent system in asymmetric information scenarios includes, for a target agent, pre-acquiring dialogue information of at least one peer target agent in the same relational sub-network, specifically including: Pre-acquire dialogue information of at least one peer target agent in the same relational subnetwork; The received dialogue information is processed by precise memory processing and fuzzy memory processing respectively to obtain precise dialogue information and summary information of the dialogue information.

[0009] According to the present invention, a collaborative method for multi-agent systems in information asymmetry scenarios involves performing precise memory processing on received dialogue information to obtain precise dialogue information, including: The received dialogue information is subjected to precise keyword matching query to obtain the precise dialogue information.

[0010] According to the present invention, a collaborative method for a multi-agent system in the case of information asymmetry is provided, wherein the received dialogue information is processed by fuzzy memory to obtain summary information of the dialogue information, including: segmenting the received dialogue information to determine at least one segment of dialogue information with a coherent context; and extracting summaries from the at least one segment of dialogue information to obtain summary information of each segment of dialogue information.

[0011] According to the present invention, a collaborative method for multi-agent systems in information asymmetry scenarios generates unknown and known inference bases based on the input task and pre-acquired dialogue information between the target agent and at least one other target agent, including: Based on the input task, a corresponding initial plan is determined, wherein the initial plan includes an initial reasoning basis for answering the question, and the initial state of the initial reasoning basis is an unknown state; Based on the initial reasoning, query the pre-acquired dialogue information between the target agent and at least one other target agent. Update the initial reasoning of the dialogue information with which the answer is found to the known reasoning, and treat the initial reasoning of the dialogue information with which the answer is not found as the unknown reasoning.

[0012] According to the present invention, a collaborative method for a multi-agent system in a situation of information asymmetry is provided, wherein the questioning includes multiple components; Based on the unknown reasoning basis, a question is generated and sent to the target agent. The corresponding unknown reasoning basis is then updated to known reasoning basis based on the response from the target agent, including: Multiple questions are generated based on the unknown reasoning basis, and a single question is sent to the target agent in each round of inquiry. The unknown reasoning basis corresponding to the single question is updated to the known reasoning basis based on the response returned by the target agent.

[0013] According to the collaborative method of a multi-agent system for information asymmetry provided by the present invention, after generating unknown and known reasoning bases, the method further includes: Receive questions sent by the target agent, retrieve the known reasoning basis, generate corresponding response content and send it to the target agent.

[0014] According to the present invention, a collaborative method for a multi-agent system in asymmetric information scenarios is provided, wherein the answer dialogue information includes precise answer dialogue information and summary answer information; Receive questions sent by the target agent, retrieve the known reasoning basis, generate corresponding response content and send it to the target agent, including: Receive questions sent by the target agent, and retrieve precise answer dialogue information and summary answer information from the known reasoning basis; Based on the retrieved precise answer dialogue information and summary answer information, a corresponding reply is generated and sent to the target agent.

[0015] According to the present invention, a collaborative method for multi-agent systems in information asymmetry scenarios involves collecting known inferences about the final update completion of the target agent and the known inferences about the final update completion of the counterparty target agent, performing consensus inference, and obtaining a final answer for responding to the task, including: Collect the updated plans of the target agent and the updated plans of the opposing target agent, wherein the updated plans include the known reasoning basis for the final update; Cross-validate the updated plans of the target agent and the opposing target agent to remove conflicting known inferences between the updated plans of the target agent and the opposing target agent. Based on the cross-validated plan of the target agent and the plan of the opposing target agent, the final answer is obtained through reasoning.

[0016] The present invention also provides a collaborative device for a multi-agent system in the case of information asymmetry, for a target agent, comprising the following modules: The task receiving module is used to receive input tasks; The dialogue information processing module is used to generate unknown reasoning basis and known reasoning basis based on the input task and the pre-acquired dialogue information between the target agent and at least one other target agent. The information interaction module is used to generate a question based on the unknown reasoning basis and send it to the target agent, and update the corresponding unknown reasoning basis to a known reasoning basis based on the reply content returned by the target agent. The consensus reasoning module is used to collect the known reasoning basis for the final update completion of the target agent and the known reasoning basis for the final update completion of the opposing target agent, perform consensus reasoning, and obtain the final answer for responding to the task.

[0017] According to the present invention, a collaborative device for a multi-agent system oriented towards information asymmetry is provided, wherein the dialogue information processing module is specifically used for: Based on the input task, a corresponding initial plan is determined, wherein the initial plan includes an initial reasoning basis for answering the question, and the initial state of the initial reasoning basis is an unknown state; Based on the initial reasoning, query the pre-acquired dialogue information between the target agent and at least one other target agent. Update the initial reasoning of the dialogue information with which the answer is found to the known reasoning, and treat the initial reasoning of the dialogue information with which the answer is not found as the unknown reasoning.

[0018] According to the present invention, a collaborative device for a multi-agent system oriented towards information asymmetry includes multiple question types; the information interaction module is specifically used for: Multiple questions are generated based on the unknown reasoning basis, and a single question is sent to the target agent in each round of inquiry. The unknown reasoning basis corresponding to the single question is updated to the known reasoning basis based on the response returned by the target agent.

[0019] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a collaborative method for a multi-agent system under any of the above-described information asymmetry conditions.

[0020] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a collaborative method for a multi-agent system under information asymmetry as described above.

[0021] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a collaborative method for a multi-agent system under information asymmetry as described above.

[0022] This invention provides a collaborative method and apparatus for multi-agent systems under information asymmetry. The method involves a target agent receiving an input task, generating unknown and known inference bases based on the input task and pre-acquired dialogue information between the target agent and at least one other target agent. Then, a question is generated based on the unknown inference base and sent to the other target agent. The unknown inference base is updated to a known inference base based on the response from the other target agent. This allows agents to proactively and dynamically acquire the latest external information and exchange and process it through effective communication strategies, improving the efficiency and accuracy of information acquisition and processing. Finally, the updated known inference bases from both the target agent and the other target agent are collected for consensus inference to obtain the final answer to the task. This solves the information asymmetry problem, ensures that each target agent can obtain all the information needed to complete the task, and achieves more efficient multi-agent collaboration and information processing capabilities. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the structure of a multi-agent system provided by the present invention.

[0025] Figure 2 This is one of the flowcharts illustrating a collaborative method for a multi-agent system under information asymmetry conditions provided by the present invention.

[0026] Figure 3 This is the second flowchart illustrating a collaborative method for a multi-agent system under information asymmetry conditions provided by the present invention.

[0027] Figure 4 This is the third flowchart illustrating a collaborative method for a multi-agent system under information asymmetry conditions provided by the present invention.

[0028] Figure 5 This is a schematic diagram of the framework of a multi-agent system provided by the present invention.

[0029] Figure 6This is a schematic diagram of a collaborative method for a multi-agent system provided by the present invention.

[0030] Figure 7 This is a schematic diagram of the collaboration between two agents provided by the present invention.

[0031] Figure 8 This is a schematic diagram of the structure of a collaborative device for a multi-agent system under information asymmetry conditions provided by the present invention.

[0032] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0034] Existing technologies in multi-agent systems (MAS) and their applications in handling information asymmetry suffer from several significant drawbacks. These drawbacks primarily affect the overall system performance, collaboration efficiency, and accuracy of information processing. The following provides a detailed description and analysis of the problems and shortcomings of existing technologies: 1) Information asymmetry problem: In existing multi-agent systems, information asymmetry is a prevalent and difficult-to-solve problem. Each agent holds different information, and this information is often complementary for completing the task. Due to the lack of effective communication and collaboration mechanisms, agents cannot fully exchange information, leading to biased decision-making and affecting the overall performance of the task.

[0035] 2) Limitations in information acquisition and processing: Existing multi-agent systems typically rely on pre-collected and stored information, which limits agents to accessing information only from their users or local environments, resulting in a limited global perspective. Because they cannot dynamically acquire and process the latest external information, agents often base their reasoning and decision-making on incomplete or outdated data, impacting the accuracy and efficiency of tasks.

[0036] 3) Privacy and security risks: While centralized information collection and sharing strategies can alleviate information asymmetry to some extent, they also bring serious privacy and security risks in practice. The centralized storage and processing of user information makes it an easy target for attacks, and it may be subject to eavesdropping and tampering during transmission.

[0037] 4) Challenges in processing dynamically changing information: Existing technologies perform poorly when processing dynamically changing information, making it difficult to adjust and optimize agent behavior and decisions in real time. The uncertainty and dynamic nature of information increases the complexity of the system, making pre-designed static models and strategies inadequate for handling complex situations in real-world applications.

[0038] 5) Insufficient system scalability and flexibility: Existing multi-agent systems are typically designed in a fixed manner, making it difficult to flexibly adjust and expand them according to specific application needs. When task scenarios or user requirements change, the system's adaptability and scalability are limited, making it difficult to meet diverse application requirements.

[0039] In summary, existing multi-agent systems have significant shortcomings in information acquisition and processing, information asymmetry, privacy and security, handling dynamic changes, and system scalability. These deficiencies limit the system's effectiveness and efficiency in complex task scenarios. The collaborative method for multi-agent systems proposed in this invention optimizes and improves upon these issues to achieve more efficient multi-agent collaboration and information processing capabilities.

[0040] Before describing the method of the embodiments of the present invention, the multi-agent system involved in the embodiments of the present invention will be described illustratively.

[0041] See Figure 1 , Figure 1 A schematic diagram of a multi-agent system is shown. It includes multiple agents forming a relational network, with each agent based on a large language model. For any target agent: the target agent and at least one peer target agent connected to it form a relational subnetwork; each target agent pre-obtains dialogue information from at least one peer target agent within the same relational subnetwork.

[0042] For example Figure 1 In the example, taking agent A as the target agent, agent A forms a relational subnetwork with its connected agents B, E, and F. The target agent can be any one of agents B, E, and F. Agent B forms a relational subnetwork with its connected agents A, C, and D. The target agent can be any one of agents A, C, and D. However, there is no connection between agent A and agents C or D, nor between agent B and agents E or F.

[0043] If target agent A and peer target agent B need to collaborate to complete task M, agent A needs to obtain its dialogue information with agents B, E, and F in advance, while agent B needs to obtain its dialogue information with agents A, C, and D in advance. It should be noted that for each agent, the dialogue information does not need to be centrally stored and managed; instead, it is distributed in each user's own storage space. Agents only retrieve the dialogue information needed to complete the task based on requests.

[0044] After target agent A and target agent B each obtain the dialogue information of at least one target agent in their respective relational subnetworks, target agent A and target agent B exchange information and cooperate to solve the problem of information asymmetry.

[0045] The following is combined with Figures 2-7 This invention describes a collaborative method for a multi-agent system under information asymmetry conditions, according to an embodiment of the present invention.

[0046] Figure 2 This is one of the flowcharts illustrating a collaborative method for multi-agent systems under information asymmetry conditions provided in this embodiment of the invention, such as... Figure 2 As shown, the method includes the following: Step 201: Receive the input task.

[0047] In a multi-agent system, the target agent can be determined based on manually input instructions.

[0048] Tasks can be input into the large language model corresponding to each target agent in the form of prompts and questions. Alternatively, tasks can be input into one target agent, which then synchronizes them to the other target agents. This embodiment uses inputting a task into one target agent as an example.

[0049] The writing process can involve two or more target proxies. In the case of multiple target proxies collaborating, it can be understood that collaboration occurs between every two target proxies. Therefore, this embodiment of the invention will use the collaboration between two target proxies as an example for explanation.

[0050] Step 202: Based on the input task and the pre-acquired dialogue information between the target agent and at least one other target agent, generate unknown reasoning basis and known reasoning basis.

[0051] In this process, the target agent pre-obtains its dialogue information with at least one other target agent. This dialogue information is generated based on the daily dialogue process between each target agent and the other target agent, and is distributed in each user's own storage space. The target agent can request and query the dialogue information needed to complete the task.

[0052] Among them, the reasoning basis is the basis used to obtain the final answer. Unknown reasoning basis is content that the target agent is unaware of when generating the answer, while known reasoning basis is content that the target agent has obtained that can be used to generate the answer.

[0053] For example, in a relational network, there are agents A, B, C, and D, connected sequentially. The current task is "Today we need to hold an online meeting, and we need to inquire about the available time slots for each agent." The selected target agents are agents B and C. For target agent B, we can directly obtain the available time of target agents A and C as known inference information, while the available time of agent D is unknown inference information. For target agent C, we can directly obtain the available time of target agents B and D as known inference information, while the available time of target agent A is unknown inference information.

[0054] Step 203: Generate a question based on the unknown reasoning basis and send it to the target agent. Update the corresponding unknown reasoning basis to a known reasoning basis based on the response returned by the target agent.

[0055] Step 203 is the step of information exchange between the target agent and the other target agent. Through information exchange between the target agent and the other target agent, the unknown reasoning basis of the target agent can be transformed into known reasoning basis, so that the information obtained by the target agent can form the answer to the task.

[0056] Continuing with the previous example, target agent B can send the question "Do you know the idle time of agent D?" to target agent C. After receiving the reply, it can determine agent D's idle time and update the unknown inference basis to a known inference basis. Alternatively, target agent B can also receive the question "Do you know the idle time of agent A?" from target agent C and generate a corresponding reply containing agent A's idle time, sending it to target agent C.

[0057] Step 204: Collect the known reasoning basis for the final update completion of the target agent and the known reasoning basis for the final update completion of the opposing target agent, perform consensus reasoning, and obtain the final answer for responding to the task.

[0058] After the target agent obtains the final updated known reasoning basis, there may be inconsistencies due to missing or incomplete information. Therefore, in step 204, consensus reasoning needs to be performed between the final updated known reasoning basis of the target agent and the final updated known reasoning basis of the opposing target agent to cross-validate their reasoning basis and obtain the final reasoning basis that can be used to generate the answer.

[0059] Specifically, step 204 includes: collecting the updated plan of the target agent and the updated plan of the opposing target agent, wherein the updated plan includes the known reasoning basis for the final update; cross-validating the updated plan of the target agent and the updated plan of the opposing target agent to remove conflicting known reasoning basis in the updated plans of the target agent and the opposing target agent; and reasoning based on the cross-validated plan of the target agent and the plan of the opposing target agent to obtain the final answer.

[0060] Consensus reasoning involves unifying the known reasoning bases of both agents, removing conflicting bases, and finally reasoning based on this unified set of known bases to arrive at the final answer. The advantage of both agents maintaining a plan containing the reasoning bases is that: their perspectives differ, and while maintaining information asymmetry, each agent updates their plan independently. Information is only merged and asymmetry removed during the final consensus reasoning, allowing for cross-verification of the correctness of the obtained reasoning bases.

[0061] This invention provides a collaborative method for multi-agent systems under information asymmetry. The method involves a target agent receiving an input task, generating unknown and known inference bases based on the input task and pre-acquired dialogue information between the target agent and at least one other target agent. Then, a question is generated based on the unknown inference base and sent to the other target agent. The corresponding unknown inference base is updated to known inference base based on the response from the other target agent. This allows agents to proactively and dynamically acquire the latest external information and exchange and process it through effective communication strategies, improving the efficiency and accuracy of information acquisition and processing. Finally, the updated known inference bases from both the target agent and the other target agent are collected for consensus inference to obtain the final answer for responding to the task. This solves the information asymmetry problem, ensures that each agent can obtain all the information needed to complete the task, and achieves more efficient multi-agent collaboration and information processing capabilities.

[0062] Furthermore, in order to effectively manage and retrieve the dialogue information of each agent, this invention designs a hybrid memory mechanism, including precise memory and fuzzy memory.

[0063] See Figure 3 For the target agent, the dialogue information of at least one peer target agent in the same relational subnetwork is obtained in advance, specifically including: Step 301: Pre-acquire the dialogue information of at least one counterpart target agent in the same relational sub-network.

[0064] Step 302: Perform precise memory processing and fuzzy memory processing on the received dialogue information to obtain precise dialogue information and summary information of the dialogue information.

[0065] Among these, precise memory stores human information in a structured manner, supporting exact match queries. It maintains the original form of information, remaining faithful to the facts without any information loss, thus ensuring accuracy. In the second step of information navigation, the target agent generates a query based on the previous round of statements from the target agent, achieving precise retrieval through a structured query language.

[0066] The received dialogue information is processed for precise memory to obtain precise dialogue information, including: performing precise keyword matching query on the received dialogue information to obtain the precise dialogue information.

[0067] Fuzzy memory stores a summary text of a conversation. A conversation refers to a series of actions over a period of time, often with a coherent context. Conversations are widely used in dialogue and recommendation systems, and there are mature extraction techniques available. The agent automatically extracts the user's conversation information, summarizes it, and stores it as fuzzy memory. Because fuzzy memory is summarized text, some details may be lost, making it somewhat fuzzy. However, fuzzy memory can store information containing longer contextual content, complementing precise memory which can only retrieve single pieces of information. The agent uses an embedding-based approximate nearest neighbor retrieval technique to retrieve fuzzy memory, achieving semantic-level information retrieval.

[0068] The received dialogue information is processed by fuzzy memory to obtain summary information of the dialogue information, including: segmenting the received dialogue information to determine at least one segment of dialogue information with a coherent context; and extracting summaries from each of the at least one segment of dialogue information to obtain summary information of each segment of dialogue information.

[0069] During each retrieval, the agent simultaneously searches both precise and fuzzy memories to ensure that the information obtained is accurate and comprehensive. This enables effective communication with the other party's agent to eliminate information asymmetry and thus resolve the problem smoothly.

[0070] After each target agent receives a task, it can first retrieve the reasoning basis for answering the question from the pre-acquired dialogue information with the other target agent, and then determine the unknown reasoning basis. For details, see... Figure 4 Step 202 includes: Step 401, determining the corresponding initial plan based on the input task, wherein the initial plan includes the initial reasoning basis for answering the question, and the initial state of the initial reasoning basis is unknown.

[0071] Step 402: Based on the initial reasoning basis, query the pre-acquired dialogue information between the target agent and at least one other target agent, update the initial reasoning basis of the queried answer dialogue information to the known reasoning basis, and treat the initial reasoning basis of the dialogue information of the unqueried answer as the unknown reasoning basis.

[0072] Steps 401-402 determine the known and unknown inference bases of the target agent. The unknown inference bases are used to generate the corresponding questions.

[0073] For cases where there are multiple questions, step 203 includes: generating multiple questions based on the unknown reasoning basis, sending a single question to the target agent in each round of inquiry, and updating the unknown reasoning basis corresponding to the single question to a known reasoning basis based on the response content returned by the target agent.

[0074] By using multiple rounds of questioning to answer multiple questions, the pressure on the agent to answer all questions at the same time can be avoided.

[0075] Specifically, after generating the unknown reasoning basis and the known reasoning basis, the method further includes: receiving a question sent by the target agent, retrieving the known reasoning basis, generating corresponding response content and sending it to the target agent.

[0076] When the target agent receives a question, it receives the question sent by the other target agent, retrieves the exact answer dialogue information and summary answer information from the known reasoning basis, and generates corresponding reply content based on the retrieved exact answer dialogue information and summary answer information and sends it to the other target agent.

[0077] By retrieving precise answer dialogue information and summary answer information each time, the target agent can ensure that the information obtained is accurate and comprehensive, so as to communicate effectively with the other agent and eliminate information asymmetry.

[0078] To facilitate understanding of the solutions in the embodiments of the present invention, a method is also disclosed in the embodiments of the present invention, and the method of the embodiments of the present invention is illustrated herein. See also Figure 5 and Figure 6 , Figure 5 The diagram shown is a framework diagram of a multi-agent system according to an embodiment of the present invention. Figure 6 The diagram shown is a flowchart of a collaborative method for a multi-agent system according to an embodiment of the present invention.

[0079] This method is illustrated using the collaboration between two agents A and B as an example. Specifically, it includes: 601. Target agent A obtains in advance the precise dialogue information M between itself and at least one other target agent. D1 Summary information M of the dialogue information F1 Target agent B obtains in advance the precise dialogue information M between the target agent and at least one other target agent. D2 Summary information M of the dialogue information F2 .

[0080] Among them, the precise dialogue information M obtained by target agent A D1 Summary information M of the dialogue information F1 The dialogue information I1 originates from its relational subnetwork; the precise dialogue information M obtained by the target agent B. D2 Summary information M of the dialogue information F2 It originates from the dialogue information I2 in its relational subnetwork.

[0081] Because the reasoning basis R required to answer task Q is distributed in different relational subnetworks, information asymmetry occurs.

[0082] 602. Input task Q into target agent A and target agent B respectively.

[0083] In this context, for target agent A, target agent B is the counterparty's target agent; and for target agent B, target agent A is the counterparty's target agent.

[0084] 603. Based on the input task, target agent A and target agent B determine the corresponding initial plans P1 and P2.

[0085] The initial plans P1 and P2 include initial reasoning for answering the question, and the initial state of the initial reasoning is unknown.

[0086] 604. Through target agent A, query the pre-acquired dialogue information I1 between the target agent and at least one opposing target agent based on the initial inference basis. Update the initial inference basis of the queried answer dialogue information to known inference basis, and treat the initial inference basis of the dialogue information without queried answer as unknown inference basis. Through target agent B, query the pre-acquired dialogue information I2 between the target agent and at least one opposing target agent based on the initial inference basis. Update the initial inference basis of the queried answer dialogue information to known inference basis, and treat the initial inference basis of the dialogue information without queried answer as unknown inference basis.

[0087] 605. Communication between target agent A and target agent B: Target agent A generates a question based on unknown reasoning basis and sends it to the other party target agent B. The unknown reasoning basis is updated to known reasoning basis according to the reply content returned by target agent B. Target agent A receives the question sent by the other party target agent B, retrieves the known reasoning basis, generates the corresponding reply content and sends it to the other party target agent B. The target agent B generates a question based on unknown reasoning and sends it to the target agent A. Based on the response from the target agent A, the unknown reasoning is updated to known reasoning. The target agent B receives the question sent by the target agent A, retrieves the known reasoning, generates the corresponding response, and sends it to the target agent A.

[0088] 606. Target agent A collects the known reasoning basis for the final update of target agent A and the known reasoning basis for the final update of the opposing target agent B to perform consensus reasoning. That is, target agent A collects the updated plans P1 and P2 of target agent A and opposing target agent B; cross-validates the updated plans P1 and P2 of target agent A and opposing target agent B, and removes conflicting known reasoning basis in the updated plans P1 and P2.

[0089] 607. Target agent A infers based on cross-validated plans P1 and P2 to obtain the final answer Ans used to respond to the task.

[0090] Of course, the execution subject of steps 606 and 607 can also be the target agent B, which outputs the final answer for the task. This embodiment will not elaborate further.

[0091] Compared with the prior art, the multi-agent system of the present invention has significant technical, economic and social benefits. Specifically, the present invention significantly improves the efficiency of information acquisition and processing by introducing an information navigation mechanism and a hybrid memory mechanism.

[0092] 1) By applying the information navigation mechanism, the multi-agent system of this embodiment of the invention achieves significant improvements in accuracy of 26.67%, 14.44%, and 4.91% respectively on three tasks ranging from simple to difficult when handling information asymmetric reasoning tasks. When handling information asymmetric localization tasks, it achieves a 12% improvement in accuracy on small relational networks and a 0.79% improvement in accuracy on large relational networks.

[0093] 2) By applying a hybrid memory mechanism, in the task of locating information asymmetric networks in large relational networks, the fuzzy memory of this embodiment of the invention brings a 6.34% improvement in accuracy, while clear memory brings a 2.38% improvement in accuracy.

[0094] The multi-agent system proposed in this invention supports agents to initiate multiple communications proactively and recursively, thereby facilitating system expansion to adapt to dynamic network changes and enabling automatic communication between agents to spread throughout the relationship network. After enabling the recursive communication function, the system achieved a 3% accuracy improvement in small-scale relationship network information location tasks and a 12.69% accuracy improvement in large-scale relationship network information location tasks.

[0095] Taking the task of locating information in a large relationship network as an example, the agent can analyze a relationship network containing 140 tasks and 588 relationship chains within 3 minutes, complete 30 rounds of communication, retrieve nearly 70,000 messages, and finally successfully complete the designated task.

[0096] To further understand the collaboration method between different agents in this embodiment, see [link to relevant documentation]. Figure 7 , Figure 7 The collaboration process between target agents A and B in this embodiment is illustrated.

[0097] In this case, target agent A obtained its dialogue information with target agents E and F in advance, and target agent B obtained its dialogue information with target agents C and D in advance.

[0098] Upon receiving the task, target agent A and target agent B each determine their corresponding initial plans. Based on the initial reasoning basis, they query the pre-acquired dialogue information between the target agent and at least one other target agent. The initial reasoning basis for the dialogue information with answers is updated to known reasoning basis, and the initial reasoning basis for the dialogue information with no answers is treated as unknown reasoning basis.

[0099] Taking the target agent B as an example, the known reasoning basis includes: its dialogue information with agent C includes summary information Session1 with C, precise dialogue information Session2 with C, and precise dialogue information Session3 with C; and its dialogue information with agent D includes precise dialogue information Session1 with D and summary information Session2 with D.

[0100] In the first round of dialogue, target agent A generates question 1, "I have no information, could you tell me…", based on unknown inference basis and sends it to target agent B. Target agent B returns "Of course. Here is…, and please tell me…". The response from target agent B includes both the reply and question 2. Target agent A updates its unknown inference basis to known inference basis based on the reply from target agent B.

[0101] In the second round of dialogue, target agent A generates a response to the opposing target agent B, and continues to generate a question "got it! and...?" based on the unknown inference basis, sending it to the opposing target agent B as well. Agent B returns "Now I know all the information. And..." Based on the response returned by target agent A, the opposing target agent B updates the unknown inference basis to the known inference basis; target agent A updates the unknown inference basis to the known inference basis based on the response returned by the opposing target agent B.

[0102] After the second round of dialogue, target agent A and the opposing target agent B cross-validate the updated plan to remove conflicting known inferences from the updated plan; they then reason based on the cross-validated plan to obtain the final answer.

[0103] Compared with the prior art, the multi-agent system provided in this invention solves the following key technical problems: 1) Information asymmetry problem: In existing technologies, due to the lack of effective communication and collaboration mechanisms, the information asymmetry problem among agents is severe, affecting the effectiveness and efficiency of task execution. This invention introduces an information navigation mechanism to guide agents to actively exchange and collaborate on information. Through plan generation and updating, the information asymmetry problem is gradually resolved, ensuring that each agent can obtain all the information needed to complete the task.

[0104] 2) Information Acquisition and Processing Efficiency Issues: In existing technologies, agents can only access information within their own users or local environments, resulting in a limited global perspective and low efficiency in information acquisition and processing. The multi-agent system proposed in this invention enables agents to proactively and dynamically acquire the latest external information and exchange and process it through effective communication strategies, thereby improving the efficiency and accuracy of information acquisition and processing.

[0105] 3) Privacy and Security Issues: Existing multi-agent systems suffer from privacy leaks and security risks during centralized information management and sharing. The multi-agent system in this invention does not require centralized information storage and management; instead, information is distributed across each user's own storage space. Agents only query the information needed to complete tasks based on requests. Furthermore, users can set different permissions for their own information. When an agent requests sensitive information, user authorization is required, thus significantly improving privacy and security compared to previous systems.

[0106] 4) Problem of handling dynamically changing information: Existing technologies struggle to cope with dynamically changing task environments and information, resulting in insufficient system flexibility and adaptability. The agent in this invention can proactively initiate communication on behalf of the user and automatically spread within the user's relationship chain, thus dynamically adapting to changes in the relationship network.

[0107] 5) System scalability and flexibility: Existing multi-agent systems have fixed designs, making it difficult to flexibly adjust and expand them according to specific application needs. The system functions of this invention can be expanded as needed, including agent communication strategies, tool usage, memory modes, etc., all of which can be dynamically adjusted and expanded, enhancing the system's scalability and flexibility and meeting diverse application requirements.

[0108] The following describes the collaborative device for a multi-agent system under information asymmetry provided by the present invention. The collaborative device for a multi-agent system under information asymmetry described below and the collaborative method for a multi-agent system under information asymmetry described above can be referred to in correspondence with each other.

[0109] This invention provides a collaborative device for a multi-agent system under information asymmetry, for a target agent. See [link to relevant documentation]. Figure 8 ,include: Task receiving module 801 is used to receive input tasks; The dialogue information processing module 802 is used to generate unknown reasoning basis and known reasoning basis based on the input task and the dialogue information between the target agent and at least one other target agent obtained in advance. The information interaction module 803 is used to generate a question based on the unknown reasoning basis and send it to the target agent, and update the corresponding unknown reasoning basis to a known reasoning basis based on the reply content returned by the target agent. The consensus reasoning module 804 is used to collect the known reasoning basis for the final update completion of the target agent and the known reasoning basis for the final update completion of the counterparty target agent to perform consensus reasoning and obtain the final answer for responding to the task.

[0110] Optionally, the apparatus further includes a preprocessing module for: Pre-acquire dialogue information of at least one peer target agent in the same relational subnetwork; The received dialogue information is processed by precise memory processing and fuzzy memory processing respectively to obtain precise dialogue information and summary information of the dialogue information.

[0111] Optionally, the preprocessing module is specifically used to: perform precise keyword matching queries on the received dialogue information to obtain the precise dialogue information.

[0112] Optionally, the preprocessing module is specifically used to: segment the received dialogue information to determine at least one segment of dialogue information with a coherent context; and extract a summary from each of the at least one segment of dialogue information to obtain a summary of each segment of dialogue information.

[0113] Optionally, the dialogue information processing module 802 is specifically used for: Based on the input task, a corresponding initial plan is determined, wherein the initial plan includes an initial reasoning basis for answering the question, and the initial state of the initial reasoning basis is an unknown state; Based on the initial reasoning, query the pre-acquired dialogue information between the target agent and at least one other target agent. Update the initial reasoning of the dialogue information with which the answer is found to the known reasoning, and treat the initial reasoning of the dialogue information with which the answer is not found as the unknown reasoning.

[0114] Optionally, the questions may include multiple items; the information interaction module 803 is specifically used for: Multiple questions are generated based on the unknown reasoning basis, and a single question is sent to the target agent in each round of inquiry. The unknown reasoning basis corresponding to the single question is updated to the known reasoning basis based on the response returned by the target agent.

[0115] Optionally, the information interaction module 803 is further configured to: after generating unknown reasoning basis and known reasoning basis, receive questions sent by the target agent, retrieve the known reasoning basis, generate corresponding reply content and send it to the target agent.

[0116] Optionally, the answer dialogue information includes precise answer dialogue information and summary answer information; The information interaction module 803 is specifically used for: Receive questions sent by the target agent, and retrieve precise answer dialogue information and summary answer information from the known reasoning basis; Based on the retrieved precise answer dialogue information and summary answer information, a corresponding reply is generated and sent to the target agent.

[0117] Optionally, the consensus inference module 804 is specifically used for: Collect the updated plans of the target agent and the updated plans of the opposing target agent, wherein the updated plans include the known reasoning basis for the final update; Cross-validate the updated plans of the target agent and the opposing target agent to remove conflicting known inferences between the updated plans of the target agent and the opposing target agent. Based on the cross-validated plan of the target agent and the plan of the opposing target agent, reasoning is performed to obtain the final answer used to respond to the task.

[0118] Figure 9 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 9 As shown, the electronic device may include a processor 910, a communications interface 920, a memory 930, and a communication bus 940, wherein the processor 910, communications interface 920, and memory 930 communicate with each other via the communication bus 940. The processor 910 can invoke logical instructions in the memory 930 to execute a collaborative method for a multi-agent system under information asymmetry. This method includes: receiving an input task; generating unknown inference criteria and known inference criteria based on the input task and pre-acquired dialogue information between the target agent and at least one other target agent; generating a question based on the unknown inference criteria and sending it to the other target agent, and updating the corresponding unknown inference criteria to known inference criteria based on the response returned by the other target agent; collecting the final updated known inference criteria of the target agent and the final updated known inference criteria of the other target agent to perform consensus inference, and obtaining a final answer for responding to the task.

[0119] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0120] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the collaborative method of a multi-agent system under information asymmetry provided by the above methods. The method includes: receiving an input task; generating unknown inference criteria and known inference criteria based on the input task and pre-acquired dialogue information between the target agent and at least one other target agent; generating a question based on the unknown inference criteria and sending it to the other target agent, and updating the corresponding unknown inference criteria to known inference criteria based on the response content returned by the other target agent; collecting the final updated known inference criteria of the target agent and the final updated known inference criteria of the other target agent to perform consensus inference to obtain the final answer for replying to the task.

[0121] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a collaborative method for a multi-agent system under information asymmetry conditions provided by the methods described above. The method includes: receiving an input task; generating unknown inference criteria and known inference criteria based on the input task and pre-acquired dialogue information between the target agent and at least one other target agent; generating a question based on the unknown inference criteria and sending it to the other target agent, and updating the corresponding unknown inference criteria to known inference criteria based on the response content returned by the other target agent; collecting the final updated known inference criteria of the target agent and the final updated known inference criteria of the other target agent to perform consensus inference to obtain a final answer for responding to the task.

[0122] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0123] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A collaborative method for multi-agent systems in situations of information asymmetry, characterized in that, The multi-agent system includes a relational network formed by multiple agents based on a large language model. For any target agent: the target agent and at least one peer target agent directly connected to it form a relational sub-network. Each target agent pre-obtains dialogue information from at least one other target agent in the same relational sub-network; The method is used for the target agent and includes: The task of receiving input; Based on the input task and the pre-acquired dialogue information between the target agent and at least one opposing target agent, generate unknown reasoning basis and known reasoning basis; Based on the unknown reasoning basis, a question is generated and sent to the target agent. The unknown reasoning basis is updated to a known reasoning basis based on the response from the target agent. Collect the known reasoning basis for the final update completion of the target agent and the known reasoning basis for the final update completion of the opposing target agent, perform consensus reasoning, and obtain the final answer for responding to the task. Based on the input task and pre-acquired dialogue information between the target agent and at least one opposing target agent, generate unknown inference bases and known inference bases, including: Based on the input task, a corresponding initial plan is determined, wherein the initial plan includes an initial reasoning basis for answering the question, and the initial state of the initial reasoning basis is an unknown state; Based on the initial reasoning basis, query the pre-acquired dialogue information between the target agent and at least one other target agent, update the initial reasoning basis of the queried answer dialogue information to the known reasoning basis, and treat the initial reasoning basis of the dialogue information of the unqueried answer as the unknown reasoning basis. Collect known reasoning evidence for the final update completion of the target agent and the known reasoning evidence for the final update completion of the opposing target agent, perform consensus reasoning, and obtain the final answer for responding to the task, including: Collect the updated plans of the target agent and the updated plans of the opposing target agent, wherein the updated plans include the known reasoning basis for the final update; Cross-validate the updated plans of the target agent and the opposing target agent to remove conflicting known inferences between the updated plans of the target agent and the opposing target agent. Based on the cross-validated plan of the target agent and the plan of the opposing target agent, reasoning is performed to obtain the final answer used to respond to the task.

2. The collaborative method for multi-agent systems in information asymmetry scenarios according to claim 1, characterized in that, For the target agent, the dialogue information of at least one peer target agent in the same relational subnetwork is obtained in advance, specifically including: Pre-acquire dialogue information of at least one peer target agent in the same relational subnetwork; The received dialogue information is processed by precise memory processing and fuzzy memory processing respectively to obtain precise dialogue information and summary information of the dialogue information.

3. The collaborative method for multi-agent systems in information asymmetry scenarios according to claim 2, characterized in that, The received dialogue information is processed for precise memorization to obtain precise dialogue information, including: The received dialogue information is subjected to precise keyword matching query to obtain the precise dialogue information.

4. The collaborative method for multi-agent systems in information asymmetry scenarios according to claim 2, characterized in that, The received dialogue information is processed using fuzzy memory to obtain a summary of the dialogue information, including: The received dialogue information is segmented to determine at least one segment of dialogue information with a coherent context. Summarize at least one segment of dialogue information to obtain summary information for each segment of dialogue information.

5. The collaborative method for multi-agent systems in information asymmetry scenarios according to claim 1, characterized in that, The questions to be asked include multiple items; Based on the unknown reasoning basis, a question is generated and sent to the target agent. The corresponding unknown reasoning basis is then updated to known reasoning basis based on the response from the target agent, including: Multiple questions are generated based on the unknown reasoning basis, and a single question is sent to the target agent in each round of inquiry. The unknown reasoning basis corresponding to the single question is updated to the known reasoning basis based on the response returned by the target agent.

6. The collaborative method for multi-agent systems in information asymmetry scenarios according to claim 1, characterized in that, After generating the unknown and known inference bases, the method further includes: Receive questions sent by the target agent, retrieve the known reasoning basis, generate corresponding response content and send it to the target agent.

7. The collaborative method for multi-agent systems in information asymmetry scenarios according to claim 6, characterized in that, The answer dialogue information includes both precise answer dialogue information and summary answer information; Receive questions sent by the target agent, retrieve the known reasoning basis, generate corresponding response content and send it to the target agent, including: Receive questions sent by the target agent, and retrieve precise answer dialogue information and summary answer information from the known reasoning basis; Based on the retrieved precise answer dialogue information and summary answer information, a corresponding reply is generated and sent to the target agent.

8. A collaborative device for multi-agent systems in situations of information asymmetry, characterized in that, The multi-agent system includes a relational network formed by multiple agents based on a large language model. For any target agent: the target agent and at least one peer target agent directly connected to it form a relational sub-network. Each target agent pre-obtains dialogue information from at least one other target agent in the same relational sub-network; The device is used for a target agent and includes: The task receiving module is used to receive input tasks; The dialogue information processing module is used to generate unknown reasoning basis and known reasoning basis based on the input task and the pre-acquired dialogue information between the target agent and at least one other target agent. The information interaction module is used to generate a question based on the unknown reasoning basis and send it to the target agent, and update the corresponding unknown reasoning basis to a known reasoning basis based on the reply content returned by the target agent. The consensus reasoning module is used to collect the known reasoning basis for the final update completion of the target agent and the known reasoning basis for the final update completion of the counterparty target agent, perform consensus reasoning, and obtain the final answer for responding to the task. The dialogue information processing module is specifically used for: Based on the input task, a corresponding initial plan is determined, wherein the initial plan includes an initial reasoning basis for answering the question, and the initial state of the initial reasoning basis is an unknown state; Based on the initial reasoning basis, query the pre-acquired dialogue information between the target agent and at least one other target agent, update the initial reasoning basis of the queried answer dialogue information to the known reasoning basis, and treat the initial reasoning basis of the dialogue information of the unqueried answer as the unknown reasoning basis. The consensus reasoning module is specifically used for: Collect the updated plans of the target agent and the updated plans of the opposing target agent, wherein the updated plans include the known reasoning basis for the final update; Cross-validate the updated plans of the target agent and the opposing target agent to remove conflicting known inferences between the updated plans of the target agent and the opposing target agent. Based on the cross-validated plan of the target agent and the plan of the opposing target agent, reasoning is performed to obtain the final answer used to respond to the task.

9. The collaborative device for multi-agent systems in information asymmetry scenarios according to claim 8, characterized in that, The questions to be asked include multiple items; The information interaction module is specifically used for: Multiple questions are generated based on the unknown reasoning basis, and a single question is sent to the target agent in each round of inquiry. The unknown reasoning basis corresponding to the single question is updated to the known reasoning basis based on the response returned by the target agent.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the collaborative method for multi-agent systems oriented towards information asymmetry as described in any one of claims 1 to 7.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the collaborative method for multi-agent systems in information asymmetry scenarios as described in any one of claims 1 to 7.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the collaborative method for multi-agent systems in information asymmetry scenarios as described in any one of claims 1 to 7.

Images