Information processing device, information processing method, and information processing program

Abstract

An information processing device (10) according to an embodiment includes a generation unit (151), a design unit (152), an update unit (153), and a provision unit (154). The generation unit (151) generates a plurality of agents, each of which is associated with a purpose, and knowledge possessed by each of the plurality of agents. The design unit (152) designs a composite task that combines tasks, by having the plurality of agents share with each other the purpose associated with each of the plurality of agents and a task for achieving that purpose. The update unit (153) updates the knowledge of each of the plurality of agents. The provision unit (154) provides a solution to the composite task on the basis of the knowledge of the plurality of agents.

Description

Information processing device, information processing method, and information processing program

[0001] This invention relates to an information processing device, an information processing method, and an information processing program.

[0002] Traditionally, techniques have been known that involve using agents equipped with large language models (LLMs) to perform tasks collaboratively.

[0003] For example, there is a known technique for building a team in which multiple specialist agents adaptively cooperate by defining the role of each agent and extracting knowledge about that role from LLM (see, for example, Non-Patent Documents 1, 2, and 3).

[0004] Furthermore, there is a known technique for dealing with complex tasks by dividing predefined tasks into individual tasks for each agent, thereby sharing the individual solutions generated by the agents within the team (see, for example, Non-Patent Documents 1 and 3).

[0005] Guangyao Chen, Siwei Dong, Yu Shu, Ge Zhang, Jaward Sesay, Boerje Karlsson, Jie Fu, and Yemin Shi. 2024a. Autoagents: A framework for automatic agent generation. In Proc. 12CAI'24, pages 22-30.Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, YaxI, Lu, Yi-Hsin Hung, Chen Qian, Yujia Qin, Xin Cong, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. 2024c. Agentverse: FacI, Litatingmulti-agent collaboration and exploring emergent behaviors. In Proc. ICLR'24.ZhenhaI, Long Wang, Shaoguang Mao, Wenshan Wu, Tao Ge, Furu Wei, and Heng Ji. 2023. Unleashing cognitive synergy in large language models: A task-solving agent through multi-person self-collaboration. CoRR, abs / 2307.05300.Qinyuan Ye, Mohammed Ahmed, Reid Pryzant, and Fereshte Khani. 2024. Prompt engineering a prompt engineer. In ACL'24 (Findings), pages 355-385. Association for Computational Linguistics.

[0006] It is a good idea to have a snack in the woods.

[0007] In conventional technologies, complex tasks assigned to teams are simply configured by combining multiple tasks, resulting in complex tasks composed of arbitrarily combined, heterogeneous tasks. However, not all agents can understand such complex complex tasks, which often leads to decreased task resolution accuracy.

[0008] Furthermore, conventional technologies lack a mechanism to utilize knowledge gained from other agents during the collaborative process, making it impossible for agents to understand the relationship between individual tasks and the team's combined tasks. As a result, agents struggle to integrate individual solutions into a comprehensive solution, and may fail to solve combined tasks accurately.

[0009] Therefore, the present invention aims to solve complex tasks with high accuracy.

[0010] To solve the problem, the information processing device of the present invention is characterized by comprising: a generation unit that generates a plurality of agents, each with an associated purpose, and knowledge for each of the plurality of agents; a design unit that designs a composite task by having the plurality of agents share the aforementioned purposes and tasks for solving those purposes with each of the plurality of agents; an update unit that updates the aforementioned knowledge for each of the plurality of agents; and a provision unit that provides a solution for the composite task based on the knowledge of the plurality of agents.

[0011] According to the present invention, complex tasks can be solved with high accuracy.

[0012] Figure 1 is a diagram showing an example configuration of an information processing device according to the first embodiment. Figure 2 is a flowchart showing the processing flow of the information processing device. Figure 3 is a diagram illustrating agent generation and team meetings. Figure 4 is a diagram illustrating Break Time and production meetings. Figure 5 is a diagram showing quantitative evaluation of task resolution accuracy between different methods. Figure 6 is a diagram showing qualitative evaluation of task resolution accuracy between different methods. Figure 7 is a diagram showing an example configuration of a computer that executes an information processing program.

[0013] The embodiments for carrying out the present invention will be described below with reference to the drawings. The present invention is not limited to these embodiments.

[0014] As mentioned above, conventional technologies that use agents to coordinate tasks have limitations, such as the inability to understand complex tasks and the inability to integrate each agent's individual solutions into a comprehensive solution.

[0015] In contrast, humans deepen and utilize the knowledge of team members through discussions both within and outside the team, thereby completing their own tasks and combined tasks. In doing so, humans improve the accuracy of their results by dynamically incorporating the latest knowledge gained from each individual's participation in these discussions.

[0016] In the first embodiment, the method used by human teams to handle complex tasks is incorporated into task resolution by a team of AI agents to improve the accuracy of solving complex tasks.

[0017] For example, suppose there is a first task, "Understand effective time management techniques," and a second task, "Learn stress management techniques." Simply linking these tasks together results in a combined task: "Understand effective time management techniques and learn stress management techniques."

[0018] Suppose we assign the first task to the first agent and the second task to the second agent. In this case, the context of the combined task may be somewhat abrupt and not sufficiently clear to the agents. Therefore, LLM may be more likely to generate a response by simply concatenating the solutions.

[0019] In contrast, designing a composite task such as "How can we explain how to reduce stress through time management and also provide independent solutions for each technique?" makes the context clearer and the relationship between individual tasks and the composite task easier to understand.

[0020] Furthermore, if the first and second agents can gain each other's knowledge through discussion, they can utilize that knowledge to formulate concrete solutions to complex task problems.

[0021] For example, the first agent can use the knowledge about "stress reduction" gained from the second agent to suggest specific solutions such as, "To manage time effectively, you should prioritize tasks and incorporate mindfulness when you have a heavy workload."

[0022] In the first embodiment, the design of complex tasks that are easy for such agents to understand and the acquisition of knowledge by the agents are realized, thereby improving the accuracy of solving complex tasks.

[0023] [Configuration of the First Embodiment] The configuration of the information processing device will be explained using Figure 1. Figure 1 is a diagram showing an example of the configuration of the information processing device according to the first embodiment.

[0024] As shown in Figure 1, the information processing device 10 includes a communication unit 11, an input unit 12, an output unit 13, a storage unit 14, and a control unit 15.

[0025] The communication unit 11 is a module for data communication with other devices. The communication unit 11 is, for example, a NIC (Network Interface Card). The input unit 12 is an interface connected to input devices such as a mouse and a keyboard. The output unit 13 is an interface connected to output devices such as a speaker and a display.

[0026] The memory unit 14 stores data, programs, etc., that are referenced when the control unit 15 performs various processes. The memory unit 14 is implemented by semiconductor memory elements such as RAM (Random Access Memory) and flash memory, or by storage devices such as hard disks and optical discs. The memory unit 14 stores model information 141.

[0027] Model information 141 consists of parameters of the language model. For example, if the language model is a neural network, then model information 141 consists of parameters such as weights and biases. The language model in the first embodiment is an LLM (Language Language Model). Furthermore, the LLM described below is constructed based on model information 141.

[0028] The control unit 15 is responsible for controlling the entire information processing device 10. The functions of the control unit 15 are realized, for example, by the CPU (Central Processing Unit) executing a program stored in the memory unit 14. The control unit 15 includes a generation unit 151, a design unit 152, an update unit 153, and a provision unit 154.

[0029] The generation unit 151 generates multiple agents. The generation unit 151 also generates knowledge for each agent. The design unit 152 designs a complex task. The update unit 153 updates the agent knowledge. The provision unit 154 provides the solution to the complex task obtained by the team of multiple agents.

[0030] In the first embodiment, the information processing device 10 provides a solution to a complex task using a method called ACTS (Knowledgeable Agents to Design and Perform Complex Tasks). ACTS will be described in detail below.

[0031] In the first embodiment, the agent is given an objective and knowledge. The agent solves a task. Specifically, the agent can be described as a system (or program) that uses LLM to obtain a solution to a task based on the objective and knowledge. Each operation of the agent described below can be realized by the information processing device 10 prompting the LLM or by the information processing device 10 acquiring the output of the LLM. Furthermore, the objective, knowledge, episodes, products, etc., may be data in the form of natural language text. Techniques for having an agent perform processing that mimics human activities such as meetings and conversations are also known (see, for example, Non-Patent Documents 1, 2, and 3).

[0032] Describe the objective, task, and knowledge. The objective is defined as "what the agent wants to incorporate into the team's output through the meeting." The objective includes a concise topic and a detailed explanation related to that topic. Agent A ｉ Objective A ｉ It can be expressed as shown in equation (1), where s ｉ This is the subject. Also, dｉ This is a detailed explanation. i is an identifier that identifies an agent.

[0033]

[0034] Agent a ｉ aims to ｉ achieve the task T ｉ，ｊ through the meeting D ｊ to conduct in-depth exploration and task resolution. Here, j indicates the number of times the meeting has been held. Agent a ｉ 's task T ｉ，ｊ is represented by the pair of its task description t ｉ，ｊ and the approach description r that the agent thinks towards its solution, as shown in equation (2). ｉ，ｊ

[0035]

[0036] The initial value of the task is t ｉ，０ = s ｉ and is prepared as such. The initial value r of the approach ｉ，０ is empty. The task is explored in depth through meetings and conversations, but the corresponding objective A ｉ is given and does not change.

[0037] Knowledge is formulated based on the following strategies. The first strategy is for the agent to accumulate knowledge about itself and the interlocutor and be able to utilize it in the design of composite tasks and production activities. Specifically, knowledge is defined in a form linked to the episodic memory regarding which agent the agent has conversed with and what knowledge has been acquired. Also, knowledge serves as a material for identifying what knowledge is necessary when the agent collaborates with other agents in task design and production activities.

[0038] The second strategy is to prepare more abstracted knowledge while summarizing episodes as specific knowledge, in order not to overly strengthen the "constraints in domain knowledge extraction" from the LLM and improve the reusability of knowledge.

[0039] ​These two strategies enable team members to leverage each other's goals and knowledge when designing and solving complex tasks, resulting in a more natural process for task design and resolution, easier task concretization, and ultimately, improved accuracy in task resolution.

[0040] In the first embodiment, knowledge is formulated based on the two strategies described above. Agent a ｉ The knowledge that he possesses is shared by L agents, including himself. ｌ The set of knowledge sets {K} for (1 ≤ l ≤ L) ｉ，ｌ} ｌ＝１ Ｌ It consists of the following, Agent A ｉ a ｌ We will explain the knowledge structure using a knowledge set for K as an example. ｉ，ｌ As shown in equation (3), knowledge K ｉ，ｌ，ｊ It is represented by a set of elements.

[0041]

[0042] Counter j is the jth meeting D ｊ This demonstrates the knowledge acquired. K ｉ，ｌ，ｊ The structure can be formulated as shown in equation (4), consisting of a summary, a keyword pair, and a set of importance levels corresponding to the pair.

[0043]

[0044] Here, M is K ｉ，ｌ，ｊ This indicates the upper limit of the number of knowledge pairs that it possesses. Summary B ｉ，ｌ，ｊ ｍ is the opponent's agent a ｌ This is a short summary of a meeting or conversation. Keyword K ｉ，ｌ，ｊ ｍ These are key keywords derived from the summary.

[0045] sc ｉ，ｌ，ｊ ｍ The knowledge pair is agent a ｉ This score indicates how important it is within the context. Furthermore, this knowledge is used by agent a ｉ But other agents a ｌ The j-th episode ε obtained during the meeting ｉ，ｌ，ｊ It is linked to ε.ｉ，ｌ，ｊ It holds information about which agent (e.g., agent al) the episode is with, and a summary of the episode, and is defined as shown in equation (5).

[0046]

[0047] Agent A ｉ K is the set of knowledge that K possesses. ｉ，ｉ is, K ｉ，ｌ It has a similar knowledge structure, is generated early, and does not possess episodic memory.

[0048] Furthermore, the summarization of the text, extraction of keywords, and calculation of importance may be performed using LLM (Language-Led Memory).

[0049] Here, we will explain the processing of each part included in the control unit 15 using Figure 2. Figure 2 is a flowchart showing the processing flow of the information processing device. Note that each process can be realized by the generation unit 151, design unit 152, update unit 153, and provision unit 154 by giving prompts to the LLM (including the LLM installed in the agent). Furthermore, in the process where each part gives prompts to operate the agent, the agent may be described as the main entity of the process.

[0050] As shown in Figure 2, first, the generation unit 151 generates multiple agents based on the objective (step S101). The generation unit 151 generates multiple agents based on the given objective A ｉ Agent A to carry out the task ｉ Generates.

[0051] Next, the generation unit 151 generates knowledge for each agent (step S102). Initial value of knowledge K ｉ，ｉ，０ It is not tied to any special episode. The generating unit 151 is {K ｉ，ｉ，０ ｍ＝１} ｍ Ｍ This generates a set of keywords that are considered useful for solving the agent's objective.

[0052] On the other hand, knowledge of other agents K ｉ，ｉ，０ The initial value of is empty. Also, the initial value of the task is T. ｉ，０ As mentioned above,ｉ，０ = s ｉ It is prepared as follows. Also, the initial value of the approach is r ｉ，０ It is empty.

[0053] Figure 3 illustrates agent generation and team meetings. As shown in Figure 3(a), for example, the generation unit 151 generates agents 21, 22, 23, and 24. Agents 21, 22, 23, and 24 are assigned Task 1, Task 2, Task 3, and Task 4, respectively. The generation unit 151 also generates knowledge 311, knowledge 321, knowledge 331, and knowledge 341 for each of agents 21, 22, 23, and 24.

[0054] Returning to Figure 2, the design department 152 conducts a team meeting and designs a composite task based on the objectives and tasks of each agent (step S103). The team meeting is an example of collaborative work.

[0055] In team meetings, each agent shares their individual objectives and corresponding tasks within the team, consolidates opinions, and designs a combined task for the team.

[0056] In the technology described in Non-Patent Document 3, the design of the complex task is performed manually. In contrast, in the first embodiment, the complex task is designed through the collaborative work of agents.

[0057] Let's describe the procedure for the team meeting. Agent A ｉ Its own purpose and task {A ｉ , T ｉ，ｊ I will express my opinion with the understanding that it may contribute to the complex tasks of the team that will be formed from now on.

[0058] After listening to each agent's expressed opinion, each agent member uses their knowledge to advise on the agent's opinion from the perspective of feasibility in designing the complex task.

[0059] The expression of opinions is achieved by having the LLM generate text. Similarly, the listening of opinions is achieved by inputting the opinion text into the LLM along with prompts.

[0060] Design Department 152 randomly selects one leader from the team members. The team leader then receives opinions and advice and designs the team's complex task G from the perspective of "feasibility". Here, Design Department 152 considers the task t of each agent. ｉ，ｊ and details of the objective d ｉ，ｊ When designing the complex task G, taking these factors into consideration, the addition of supplementary information for combination, rather than simple linking, is permitted to refine the overall structure.

[0061] The design of the composite task G is expressed as shown in equation (6) using the function fuse for designing composite tasks.

[0062]

[0063] Here, the function `fuse` is implemented as an LLM prompt. `fuse` combines the details of each agent's task and the objectives they refer to, as well as supplementary information generated to integrate the tasks, to form a feasible, team-wide composite task G.

[0064] Furthermore, when combined tasks become feasible, the solutions to each agent's own tasks become easier to concretize, which in turn leads to an improvement in the accuracy of solving combined tasks.

[0065] As shown in Figure 3(b), the design unit 152 has agents 21, 22, 23, and 24 conduct a team meeting to design a composite task. The design unit 152 designs a composite task by having the multiple agents share with each other the objectives associated with each of the multiple agents, and the tasks to solve those objectives.

[0066] Returning to Figure 2, the update unit 153 updates the knowledge of each agent based on the combined task, the results of the team meeting, and each agent's objectives, tasks, and knowledge (step S104). The knowledge update involves two steps: task in-depth analysis and knowledge acquisition.

[0067] First, the update unit 153 performs task in-depth analysis. The procedure for task in-depth analysis is described below. Task in-depth analysis is performed by agent a ｉ T ｉ，ｊ , purpose A ｉ Meeting D ｊ , knowledge K ｉ，ｉ This is carried out by equation (7), assuming a complex task G.

[0068]

[0069] The function f() is implemented by prompting the LLM. The function f() prompts the agent to decide how to update its task in line with the agent's own objectives and the composite task, based on the meeting minutes. Task T is updated by the function f(). ｉ，ｊ Task T ｉ，ｊ＋１ It will be updated. Note that the number of pieces of knowledge given to the function f() is the importance score sc ｉ，ｉ The order can be c items (for example, c = 10).

[0070] Next, the update unit 153 performs knowledge acquisition. The procedure for knowledge acquisition is described below. First, agent a ｉ Meeting D ｊ After the meeting, refer to the meeting minutes and the speaker (Agent A) ｌ Based on the statement by ) Knowledge K ｉ，ｌ，ｊ And the episode that connects it ε ｉ，ｌ，ｊ To obtain this. The meeting minutes may be a history of the output of the participants (agents).

[0071] Specifically, agent a ｉ This is Conference D ｊ Previous knowledge of both agents K ｉ，ｉ _K ｉ，ｌ Meeting D ｊ Previous a ｉ Task T ｉ，ｊ，１ , purpose A ｉ Based on this, knowledge is acquired using equation (8).

[0072]

[0073] The function g() is implemented as a prompt to the LLM. When acquiring the knowledge K of the opponent (agent a ｌ ), g() focuses on what knowledge the opponent has and whether that knowledge is useful for a's own task solution or the team's composite task solution, and assigns a score sc ｉ，ｌ，ｊ to each knowledge pair (K ｉ and K ｉ，ｌ，ｊ ｍ ) while acquiring knowledge. ｉ，ｉ and K ｉ，ｌ ) while acquiring knowledge.

[0074] Note that the number of knowledge (K ｉ，ｉ and K ｉ，ｌ ) of the two agents to be given is set to c in the order of the important scores sc ｉ，ｉ or sc ｉ，ｊ .

[0075] Agent a ｉ integrates the acquired knowledge with the knowledge acquired in the past for each agent a ｌ and updates the knowledge set K ｌ for a ｌ . When doing so, agent a ｉ，ｌ integrates the vocabulary with the same meaning as the one registered in the past as a keyword. For example, if agent a ｉ has registered the keyword "automobile" in the past, when registering "car" newly, it unifies the vocabulary as the same keyword. This suppresses the context size when referring to knowledge. Note that although the knowledge is integrated, the number of knowledge pairs held by K ｉ exists as the number created in all past episodes. This is based on the premise that the summary B ｉ，ｌ，ｊ of the episodes constituting the knowledge pair and the episode ε ｉ，ｌ，ｊ and the episode ε ｉ，ｌ，ｊ are all different and cannot be summarized.

[0076] Subsequently, the update unit 153 performs Break Time using the domain expert agent (step S105). The procedure of Break Time is described below.

[0077] Each agent converses with team members and integrates tasks during team meetings. However, in order to solve both their own tasks and the complex tasks designed by the team simultaneously, a consistent organization of knowledge is necessary that satisfies both individual task resolution and the resolution of the complex tasks.

[0078] Therefore, in the first embodiment, a domain expert agent capable of answering such questions (for example, questions that clarify consistent knowledge that satisfies both the agent's own task solving and the solving of a composite task) is dynamically generated after the team meeting, and a one-on-one discussion is held with the agent. The discussion procedure is as follows.

[0079] First, agent a ｉ Agent A reflects on whether their own objectives and tasks contributed to the design of the complex task. ｉ This involves reflecting on what areas of knowledge are lacking in order to concretize one's own tasks, solve complex tasks, and contribute to the team. Then, agent a ｉ This involves determining the set of knowledge you want to deepen your understanding of, and at the same time, determining the questions you want to ask about that knowledge.

[0080] Next, agent a ｉ This refers to a field expert agent who can answer the determined question. ｄ It dynamically generates the "knowledge to deepen" a ｄ Objective A ｄ For example, field expert agent a, whose initial knowledge was set using the same procedure as in steps S101 and S102. ｄ Generates.

[0081] Next, Agent A ｉ This involves asking questions about the knowledge you want to deepen (improve) to a specialist agent in that field. ｄ I ask them.

[0082] Specialist Agent A ｄ Agent A answers the question by utilizing their own knowledge. ｉThen, consider if there are any further questions you may have, and if so, ask follow-up questions.

[0083] Update unit 153 is field expert agent a ｄ Response by agent a ｉ The questions are repeatedly asked by agent a. ｉ The process continues until the participant has run out of questions to ask or until the number of utterance / response turns (the number of question-and-answer repetitions) exceeds a certain value.

[0084] Subsequently, the update unit 153 updates the knowledge of each agent based on the Break Time results (step S106). The procedure for updating the knowledge in step S106 may be the same as in step S104. However, agent a ｌ is a specialist agent in the field. ｄ It can be replaced with this.

[0085] Figure 4 illustrates Break Time and production meetings. As shown in Figure 4(c), the update unit 153 causes agent 21 and field expert agent 25 to conduct Break Time (repeated questioning and answering about knowledge). Field expert agent 25 possesses knowledge 351. The update unit 153 also updates knowledge 311 to knowledge 312 based on the results of Break Time.

[0086] Returning to Figure 2, the supply unit 154 conducts a production meeting and provides a solution to the complex task (step S107). The procedure for the production meeting is described below.

[0087] The supply unit 154 produces products that include the solutions to the complex tasks designed in the team meeting, as determined by the production meeting. The procedure for this production activity is as follows. Note that producing products that include the solutions to complex tasks is also described in Non-Patent Document 3.

[0088] Each agent a ｉ This is based on the knowledge accumulated by the individual (updated knowledge) and the knowledge of the team members {K ｉ，ｌ} ｌ＝１ Ｌ and its own task T ｉ，ｊ Refer to your objective Aｉ A concrete approach to r ｉ，ｊ To elaborate on this as a solution and incorporate it into the product, and to assign objective A to the team's combined task G. ｉ With the understanding that this aligns with and contributes to the community, each opinion R ｉ Agent A declares: ｉ Team members' knowledge from their perspective {K ｉ，ｌ} ｌ＝１ Ｌ Agent A, while maintaining consistency with the objectives of other agents, ｉ Refer to the solution to the objective in order to incorporate it into the product.

[0089] Agent a in the production meeting ｉ The expression of this opinion is represented as shown in equation (9).

[0090]

[0091] Here, the function h() is implemented as a prompt to the LLM. The function h() prompts the agent to express an opinion in a manner that aligns with the agent's own task and objective, as well as with the composite task.

[0092] The number of pieces of knowledge to be given will be c, ordered by importance score (sc).

[0093] Then, the supply unit 154 randomly selects one leader from the team members. The leader agent creates a prototype of the product, taking into account the opinions of each agent. At that time, the leader considers each agent's objective A. ｉ The following points should be confirmed: that the product contains answers to the questions, that the product fully achieves the team's objectives, and that clarity and word count are in accordance with the regulations.

[0094] The leader determines the final product once they have confirmed that the conditions (verification items) are met. If the conditions (verification items) are not met, the leader identifies the problems and shares them with the team.

[0095] Each agent, upon receiving the problem, reiterates their opinion. The supply unit 154 carries out the process from the leader's decision to the re-expression of opinions until the conditions are met. Once the conditions are met, the supply unit 154 provides (outputs) the product as a solution to the complex task.

[0096] As shown in Figure 4(d), the providing unit 154 provides a solution to the complex task based on the knowledge 312 of agent 21, the knowledge 322 of agent 22, the knowledge 332 of agent 23, and the knowledge 342 of agent 24. Knowledge 312, knowledge 322, knowledge 332, and knowledge 342 are updated versions of knowledge 311, knowledge 321, knowledge 331, and knowledge 341, respectively.

[0097] [Experiment] First, we will describe the experiment conducted to evaluate the embodiment (ACTS).

[0098] Human objectives in real society are diverse, and so are the solutions. To conduct a verification that reflects real-world situations, the experiment used a dataset with diverse objectives and solutions.

[0099] In the experiment, we constructed a dataset with diverse objectives and solutions from a non-factoid QA dataset, where the answers to the subject are not unique. Here, we designed the agent's objective from the subject and set the answers as solutions to the agent's objective. This allows us to quantitatively verify that a team designs and solves complex tasks, even though individual agents have diverse objectives and potential solutions, rather than a task that does not require consensus to include a unique answer in the output. Furthermore, the dataset is divided into five community categories: tea, coffee, design, architecture, and fashion.

[0100] The quantitative evaluation metric used in the experiment is Rouge. Rouge is a metric that measures the proportion of referenced answer sentences included in the generated text.

[0101] In the experiment, the team will consist of four agents. Since there is a correct answer for each agent, the average Rouge value measured for each agent will be used as the evaluation metric.

[0102] The dataset used in this study contains multiple candidate solutions as correct answers. To prevent variations in the number of solutions for a given objective from affecting accuracy, Rouge is calculated for each solution to the objective, and the solution with the highest Rouge is used as the accuracy representing the correctness of the answer to that objective. The generated product is a marketing plan.

[0103] Furthermore, as a qualitative evaluation, a scoring system consisting of a 0-10 scale was used to measure whether the agent's objectives were qualitatively included in the proposal, using a prescribed method. The validity of this scoring system was then cross-referenced with two experts in this field.

[0104] The comparison methods are Camel, SPP, AgentVerse (Non-Patent Document 2), and AutoAgents (Non-Patent Document 1).

[0105] Figures 5 and 6 show the experimental results. Figure 5 shows the quantitative evaluation of task resolution accuracy between different methods. Figure 6 shows the qualitative evaluation of task resolution accuracy between different methods.

[0106] [Effects of the First Embodiment] As described above, the generation unit 151 generates multiple agents, each with an associated purpose, and the knowledge of each of the multiple agents. The design unit 152 designs a composite task by having the multiple agents share the purposes associated with each of the multiple agents and the tasks to solve those purposes with each other. The update unit 153 updates the knowledge of each of the multiple agents. The provision unit 154 provides a solution to the composite task based on the knowledge of the multiple agents.

[0107] As a result, the information processing device 10 can integrate tasks in a manner that makes the context clearer and the relationship between individual tasks and the combined task easier to understand, rather than simply combining multiple tasks. Furthermore, the information processing device 10 can enable the agent to understand the relationship between individual tasks and the team's combined task. Consequently, according to the first embodiment, the combined task can be solved with high accuracy.

[0108] Furthermore, the design unit 152 designs a complex task by having each of the multiple agents output their own objectives and tasks, as well as their opinions on how their tasks contribute to the team's task, and providing advice on those opinions. This allows the information processing device 10 to automatically design complex tasks without manual intervention.

[0109] The update unit 153 updates the knowledge of the first agent based on the results of repeatedly receiving questions from the first agent among multiple agents and answers from domain expert agents who can answer both questions related to the first agent's task and questions related to composite tasks.

[0110] As a result, the information processing device 10 can automatically perform complex task design without human intervention. The information processing device 10 can enable the agent to organize knowledge in a consistent manner that satisfies both its own task resolution and the resolution of complex tasks.

[0111] [System Configuration, etc.] Furthermore, the components of each part shown in the diagram are functional concepts and do not necessarily need to be physically configured as shown. In other words, the specific forms of distribution and integration of each device are not limited to those shown in the diagram, and all or part of them can be functionally or physically distributed and integrated in any unit according to various loads and usage conditions. In addition, all or any part of the processing functions performed by each device can be realized by a CPU and the program executed on that CPU, or by hardware using wired logic.

[0112] Furthermore, among the processes described in the embodiments described above, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by known methods. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.

[0113] [Program] The information processing device 10 described above can be implemented by installing a program (information processing program) as packaged software or online software on a desired computer. For example, by having the computer run the above program, the computer can function as the information processing device 10. The term "computer" here includes mobile communication terminals such as smartphones, mobile phones and PHS (Personal Handyphone System), as well as terminals such as PDA (Personal Digital Assistant).

[0114] Figure 7 shows an example configuration of a computer that executes an information processing program. Computer 1000 has, for example, memory 1010 and CPU 1020. Computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0115] Memory 1010 includes ROM (Read Only Memory) 1011 and RAM (Random Access Memory) 1012. ROM 1011 stores, for example, a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090. The disk drive interface 1040 is connected to the disk drive 1100. For example, a removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, a display 1130.

[0116] The hard disk drive 1090 stores, for example, the OS 1091, application programs 1092, program modules 1093, and program data 1094. That is, the programs that define each process executed by the information processing device 10 are implemented as program modules 1093 in which executable code for a computer is written. The program modules 1093 are stored, for example, in the hard disk drive 1090. For example, a program module 1093 for executing processes similar to the functional configuration of the information processing device 10 is stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

[0117] Furthermore, the data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, memory 1010 or hard disk drive 1090. The CPU 1020 then reads the program module 1093 and program data 1094 stored in memory 1010 or hard disk drive 1090 into RAM 1012 as needed and executes them.

[0118] Furthermore, the program module 1093 and program data 1094 are not limited to being stored in the hard disk drive 1090; for example, they may be stored in a removable storage medium and read by the CPU 1020 via a disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via a network interface 1070.

[0119] 10 Information Processing Unit 11 Communication Unit 12 Input Unit 13 Output Unit 14 Storage Unit 15 Control Unit 141 Model Information 151 Generation Unit 152 Design Unit 153 Update Unit 154 Provision Unit

Claims

1. An information processing device comprising: a generation unit that generates knowledge for each of the multiple agents, each with an associated purpose; a design unit that designs a composite task by having the multiple agents share with each other the purposes and tasks for solving those purposes, each of the multiple agents; an update unit that updates the knowledge for each of the multiple agents; and a provision unit that provides a solution for the composite task based on the knowledge of the multiple agents.

2. The information processing apparatus according to claim 1, characterized in that the design unit designs the composite task by causing each of the plurality of agents to output an opinion on its own purpose and the contribution of its task to the task of the team consisting of the plurality of agents, and advice on the opinion.

3. The information processing device according to claim 1, wherein the update unit updates the knowledge of the first agent based on the results of repeatedly receiving questions from the first agent among the plurality of agents and answers to those questions from a field expert agent capable of answering both the questions concerning the first agent's task and the questions concerning the composite task.

4. An information processing method performed by a computer, comprising: a generation step of generating a plurality of agents, each with an associated purpose, and knowledge for each of the plurality of agents; a design step of designing a composite task by having the plurality of agents share with each of the plurality of agents the purpose and tasks for solving the purpose associated with each of the plurality of agents; an update step of updating the knowledge for each of the plurality of agents; and a provision step of providing a solution for the composite task based on the knowledge of the plurality of agents.

5. An information processing program characterized by causing a computer to execute: a generation step of generating a plurality of agents, each with an associated purpose, and knowledge for each of the plurality of agents; a design step of designing a composite task by having the plurality of agents share with each of the plurality of agents the purpose and the task for solving the purpose associated with each of the plurality of agents; an update step of updating the knowledge for each of the plurality of agents; and a provision step of providing a solution for the composite task based on the knowledge of the plurality of agents.

Images