Information processing systems, information processing methods, and programs

Abstract

This invention provides an information processing system that uses multiple AI agents, enabling each AI agent to perform a consistent process. [Solution] An information processing system using a multi-AI agent in which multiple agents cooperate to perform tasks, wherein the multiple agents define and share a project specification document.

Description

Technical Field

[0001] The present invention relates to an information processing system, an information processing method, and a program.

Background Art

[0002] For improving a company's productivity and competitiveness, data-driven decision-making is important. However, a serious shortage of data personnel greatly hinders the utilization of data by Japanese companies. In particular, in small and medium-sized enterprises, there are many cases where employees who are not specialized in data analysis are in charge of analysis work. As a result, problems such as "not knowing how to utilize data" have occurred.

[0003] In addition, since it costs a high amount to train and hire data personnel, it is not easy to increase their numbers. There is also a prediction that Japan will be short of up to 790,000 IT personnel including data personnel by 2030, and it is considered that the shortage of data personnel will become even more serious in the future. Under such circumstances, it is an urgent task to create an environment in which all employees, not limited to data personnel, can effectively utilize data in corporate activities.

[0004] In data analysis, clarifying causal relationships contributes to accurate decision-making. In recent years, analysis using machine learning techniques has achieved high prediction accuracy, but its internal structure is regarded as a "black box" that is difficult to understand, and there are problems in its application in fields where important decision-making such as in medicine and finance is required. To solve this, there is a shift towards "Explainable AI (XAI)" that can explain the operation of the model and the basis of the results, but these methods are still analyses based on the correlation between data and are insufficient for accurately identifying causal relationships. With only correlation, the cause and result cannot be accurately grasped, and there is a risk of leading to incorrect conclusions and decisions. Therefore, for making important decisions, the introduction of an analysis method for clarifying causal relationships is highly regarded.

[0005] However, causal inference methods used to uncover cause-and-effect relationships require advanced expertise, making them difficult for non-experts to utilize. Causal inference methods include "causal exploration methods" that discover cause-and-effect relationships from data, and "effect verification methods" that examine the impact of specific measures on results. However, accurately applying these methods requires a deep understanding of statistics and econometrics, appropriate model selection, and verification of data assumptions. Therefore, given the current shortage of data professionals, it is not realistic for non-experts to perform these advanced causal analyses.

[0006] Patent Document 1 describes a system that provides means for managing the progress of a chat regarding data analysis conducted with a user, means for identifying the user's analysis objective through the progress of the chat, means for analyzing a target dataset according to the analysis objective, and means for outputting the results of the analysis of the target dataset, thereby enabling the system to obtain appropriate data analysis results regardless of the user's level of data analysis skills. [Prior art documents] [Patent Documents]

[0007] [Patent Document 1] Patent No. 7369320 [Overview of the project] [Problems that the invention aims to solve]

[0008] In the system described in Patent Document 1, data analysis required by the user is carried out via a chat-based communication between the user and the data analysis agent, and the data analysis results are ultimately obtained. However, the specific configuration of the data analysis agent is not clear. In particular, Patent Document 1 does not disclose how each AI agent cooperates to perform tasks when using multiple AI agents as the data analysis agent.

[0009] The present invention aims to provide an information processing system that uses multiple AI agents, in which each AI agent can perform a consistent process. [Means for solving the problem]

[0010] This information processing system uses a multi-AI agent where multiple agents collaborate to perform tasks, and the multiple agents define and share a project specification document. [Effects of the Invention]

[0011] In an information processing system using a multi-AI agent where multiple agents collaborate to perform tasks, a consistent process can be achieved by defining and sharing a project specification document among the multiple agents. [Brief explanation of the drawing]

[0012] [Figure 1] This diagram shows the configuration of the data analysis system according to Embodiment 1. [Figure 2] This figure shows the hardware configuration of the data analysis device according to Embodiment 1. [Figure 3] This figure shows the functional block of the arithmetic unit according to Embodiment 1. [Figure 4] This diagram shows the overall sequence of the data analysis system according to Embodiment 1. [Figure 5] This figure shows a specific example of the detailed analysis specification document related to Embodiment 1. [Figure 6] This diagram shows the processing flow of the consulting agent according to Embodiment 1. [Figure 7A] This figure shows the content of the interaction between the user and the system in the consulting agent according to Embodiment 1. [Figure 7B] This figure shows the content of the interaction between the user and the system in the consulting agent according to Embodiment 1. [Figure 8A]It is a diagram showing the processing flow of the data understanding agent according to Embodiment 1. [Figure 8B] It is a diagram showing the processing flow of the data understanding agent according to Embodiment 1. [Figure 9] It is a diagram showing the dialogue content between the user and the system in the data understanding agent according to Embodiment 1. [Figure 10] It is a diagram showing the processing flow of the analysis agent according to Embodiment 1. [Figure 11] It is a diagram showing the dialogue content between the user and the system in the analysis agent according to Embodiment 1. [Figure 12] It is a diagram showing the processing flow of the result interpretation agent according to Embodiment 1. [Figure 13] It is a diagram showing the dialogue content between the user and the system in the result interpretation agent according to Embodiment 1. [Figure 14] It is a diagram showing the processing flow of the document creation agent according to Embodiment 1. [Figure 15A] It is a diagram showing the dialogue content between the user and the system in the document creation agent according to Embodiment 1. [Figure 15B] It is a diagram showing the dialogue content between the user and the system in the document creation agent according to Embodiment 1. [Figure 16] It is a diagram showing the functional blocks of the arithmetic unit according to Embodiment 2.

Mode for Carrying Out the Invention

[0013] The system of the present invention Each AI agent inputs information, and this information is then used as a template for other AI agents to refer to (hereinafter, fills in what is called a "project specification sharing document" 。) through the dialogue between the user and the AI agent, and each AI agent acts based on the project specification sharing document to perform a consistent process, and can be widely applied to information processing systems using multi-AI agents such as data analysis, product development processes, inspection / test implementation in factories, personnel recruitment processes, and marketing processes.

[0014] As a preferred embodiment of the present invention, Embodiment 1, a data analysis process utilizing a multi-AI agent will be described below.

[0015] In Embodiment 1, a system is provided that utilizes multiple AI agents, with each AI agent performing complex tasks (requiring specialized knowledge), allowing users to execute a series of data analysis processes through chat-based dialogue with the AI ​​agents without requiring specialized knowledge. This makes it possible for non-experts to perform processes that previously required time and effort from data scientists and other specialists in a short amount of time.

[0016] In Embodiment 1, users can consistently perform tasks from defining business challenges to preparing data, analyzing results, formulating strategies, and generating reports through dialogue with an AI agent. This supports a wide range of analyses, including machine learning and causal inference techniques, enabling advanced data analysis even without specialized knowledge.

[0017] Embodiment 1 addresses the challenges of a shortage of data professionals and the difficulty of data analysis by non-experts, creating an environment where all employees can make data-driven decisions. This is expected to improve corporate productivity and strengthen competitiveness, and significantly contribute to the democratization of data science.

[0018] [System Configuration] Figure 1 shows the configuration of the data analysis system 1 according to Embodiment 1. As shown in Figure 1, in the data analysis system 1, the data analysis device 10, the administrator terminal device 20, and the user terminal device 30 are connected via the network 40.

[0019] The data analysis device 10 is a general-purpose information processing device that can be used as a dedicated system, or as a general-purpose computer, server, or multiple computers distributed on a network (cloud computer), and can implement data analysis functions by installing a program.

[0020] The administrator terminal device 20 is a device used by the administrator, i.e., the provider of the data analysis system 1, and can be any information processing device such as a general-purpose computer.

[0021] The user terminal device 30 is a device used by users of the data analysis system 1, and can be any information processing device such as a general-purpose computer. The number of user terminal devices 30 connected to the network 40 is determined according to the number of users of the data analysis system 1.

[0022] Network 40 is responsible for interconnecting the data analysis device 10, the administrator terminal device 20, and the user terminal device 30, and can use any wired or wireless network.

[0023] [Hardware configuration] Figure 2 is a diagram showing the hardware configuration of a data analysis device 10 according to Embodiment 1 of the present invention. As shown in Figure 2, the data analysis device 10 includes, as an example, an interconnected arithmetic unit 110, a storage device 120, an auxiliary storage device 130, an input device 140, an output device 150, and a communication device 160.

[0024] The arithmetic unit 110 has a processor such as a CPU that can execute an instruction set.

[0025] The storage device 120 has memory such as RAM for storing the instruction set.

[0026] The auxiliary storage device 130 has a recording medium such as an HDD or SSD for recording programs. The auxiliary storage device 130 may also be cloud storage.

[0027] The input device 140 is a device that accepts input from an external source, such as a keyboard or a mouse.

[0028] The output device 150 is a device that outputs to an external source, such as a display or a printer.

[0029] The communication device 160 has a communication interface for sending and receiving various types of data over a network.

[0030] [Function Block] Figure 3 shows the functional blocks of the computing device 110 included in the data analysis device 10 according to Embodiment 1 of the present invention. As shown in Figure 3, the computing device 110 includes a consulting agent 200, a data understanding agent 300, an analysis agent 400, a result interpretation agent 500, and a document creation agent 600.

[0031] In describing the system according to Embodiment 1 of the present invention, we will first explain the processes of the multi-AI agent and data scientist that form its premise.

[0032] [Multi-AI Agent] Multi-AI agent systems represent a new AI paradigm made possible by the emergence of Large Language Models (LLMs), where multiple AI agents collaborate to perform tasks. Here, an AI agent is defined as an entity that perceives its surroundings, makes judgments and decisions, and executes actions. While AI development has progressed from rule-based systems to machine learning and deep learning, the advent of LLMs has dramatically improved natural language comprehension and generation capabilities. By leveraging LLMs' advanced natural language processing capabilities and versatility to handle diverse tasks, it is expected that communication between AI agents and with users will become easier, enabling the tackling of complex problems.

[0033] While research into multi-agent systems had been ongoing even before the advent of LLM, their applications were limited. Conventional agents were primarily rule-based or based on specific algorithms, making them suitable for task execution in specific environments and conditions. However, they lacked flexibility and versatility, and inter-agent communication was limited, making complex information exchange and cooperation difficult. As a result, the practical application of multi-agent systems remained confined to specific domains.

[0034] However, the use of LLM has evolved multi-agent systems to a new level. LLM possesses advanced natural language processing capabilities and versatility to handle diverse tasks, enabling natural language communication between agents and with users. As a result, each agent, while possessing specialized knowledge and roles, can share information and collaborate to perform tasks through natural language. For example, in project management, a planning agent, a resource management agent, and a risk assessment agent can work together to automatically generate an optimal project plan. In this way, multi-AI agents utilizing LLM represent a new AI paradigm that can address complex problems that were difficult to handle with conventional systems.

[0035] The system according to Embodiment 1 of the present invention is a system in which such multi-AI agents take on the role of data scientists. Each AI agent is responsible for and collaboratively performs specialized data science tasks such as data collection and preprocessing, selection of analysis methods, interpretation of results, and proposal of measures. As a result, users can perform advanced data analysis through the multi-AI agents even without specialized knowledge. The system according to Embodiment 1 of the present invention, as a new application example of multi-AI agents, solves the problem of a shortage of data personnel and supports data-driven decision-making.

[0036] [The Data Scientist Process] To solve business challenges, data scientists need to follow a systematic process. This process includes everything from initial interviews and data preparation to analysis, model building, results utilization and implementation, and finally, reporting. At each step, the data scientist's skills—including business acumen, data engineering, and data science—play a crucial role.

[0037] First, the initial consultation involves analyzing the current situation and defining business challenges. This stage requires strong business acumen. Through dialogue with clients and stakeholders, a deep understanding of the organization's current state and problems is crucial, clarifying the issues that need to be addressed. Accurately grasping the business background and objectives lays the foundation for formulating an appropriate analytical approach.

[0038] Next, data preparation involves data collection, preprocessing, and trend analysis. This is where data engineering skills come into play. Necessary data is collected from diverse data sources, and preprocessing is performed, such as data cleansing and handling missing values. Exploratory data analysis is also conducted to understand the basic trends and characteristics of the data, preparing it for subsequent analysis.

[0039] The analysis, model building, and accuracy improvement steps require expertise in data science. Statistical methods and machine learning algorithms are used to extract useful insights from the data. After model construction, the model's performance is validated based on evaluation metrics, and accuracy is improved through hyperparameter tuning and feature engineering.

[0040] Next, in utilizing and operating the analysis results and models, the integration of business acumen and data science is crucial. The obtained analysis results must be interpreted within a business context and linked to concrete measures and decision-making. When deploying the model to a real-world environment and integrating it into business processes, continuous monitoring and maintenance are also necessary.

[0041] Finally, reporting requires effectively communicating the results of the analysis to stakeholders. Leveraging business skills, this involves presenting complex information in a way that is easily understandable to non-experts, thereby supporting data-driven decision-making. Visualization tools and presentation skills are used to maximize the communication of the value of the analysis results.

[0042] In the system according to Embodiment 1 of the present invention, this process is handled by multiple AI agents. Each AI agent takes on the role of a data scientist and works collaboratively to perform tasks, allowing users to perform a series of data analysis processes—from defining business problems to preparing, analyzing, interpreting results, deciding on measures, and creating reports—through dialogue with AI agents in a chat format, even without specialized knowledge like that of a data scientist. This enables them to perform advanced data analysis.

[0043] Here, the correspondence between the tasks of each AI agent and the processes of the data scientist in the system according to Embodiment 1 of the present invention is as follows. Consulting agent → Interview Data understanding agent → Data preparation Analysis agent → Analysis, model building, and accuracy improvement Results interpretation agent → Utilization and operation of analysis results and models Document creation agent → Reporting

[0044] Figure 4 shows the processing flow by each AI agent in the system according to Embodiment 1 of the present invention. Since Embodiment 1 relates to data analysis, the "Project Specification Sharing Document" will be referred to as the "Analysis Detailed Specification Document" here.

[0045] In the system of the present invention, as shown in Figure 4, a "Detailed Analysis Specification" is defined and shared as a common core that connects each task (AI agent). In other words, in the system of the present invention, each AI agent constituting the multi-AI agent engages in repeated dialogue with the user in its own task. The generated information, Items in the "Detailed Analysis Specification" In the column sequential Enter Then, the "Detailed Analysis Specification" is handed over to another AI agent, and the AI ​​agent that receives the handover then uses the "Detailed Analysis Specification". Refer to the information entered in the item field, This system allows users to perform tasks while understanding and interpreting the content.

[0046] The "Detailed Analysis Specification" of Embodiment 1 includes, for example, the following content for each category:

[0047] [basic information] This information includes user attributes obtained from interactions with users, as well as information indicating the direction of the analysis, such as the theme and objectives of the analysis.

[0048] [Analysis method] This information provides details about analytical perspectives, such as target indicators and constraints, as well as analytical methods, including expected outputs and anticipated actions agreed upon with users.

[0049] [Data used] This information summarizes the content, source, and preprocessing requirements (missing data handling, normalization, categorization, and time series processing) of the data used for analysis.

[0050] [Analysis results] This document records the results, such as graphs, obtained by an AI agent performing analysis, along with an interpretation of the analysis results from a business perspective, and the insights derived from that interpretation.

[0051] [Report] This analytical report is generated from the objectives and analysis results derived through interactions with users. The report includes the background and objectives of the data analysis, a summary of the results, and the insights and next actions derived from them.

[0052] Figure 5 shows a concrete example of the structure of the "Detailed Analysis Specification." Each AI agent performs its own task by referring to the "Detailed Analysis Specification," and the AI ​​agent listed in the "Inputting Agent" column fills in the information for each category (item) and shares that information with other AI agents.

[0053] The following provides a detailed explanation of the processing flow for each AI agent.

[0054] [Consulting Agent] Figure 6 shows the processing flow at the consulting agent.

[0055] Starting from the starting node, the process first asks the user "What do you want to analyze (analysis target)" and "What is the purpose of the analysis, such as prediction (analysis objective)?" and requests input (S100). Then, it checks whether the information entered by the user is sufficient (S110). If it is insufficient (e.g., only the analysis target is entered, or the analysis objective is unclear), it prompts the user to enter the information again, providing input examples as needed. Once it is determined that the information entered by the user is sufficient, the process proceeds to the next step.

[0056] Next, based on the information obtained from users and the database information, we consider what methods to use and decide on the analysis method (e.g., time series forecasting, classification, regression, causal exploration, effectiveness verification) (S120).

[0057] Next, it is determined whether the chosen analysis method is "time series forecasting" or not (S130). If it is "time series forecasting," additional confirmations (interviews) are conducted with the user to determine the granularity of the forecast (daily, monthly, etc.), the period of data to be used for training, and the necessary aggregation axes (S140). Aggregation axes refer to aggregation units such as product categories, departments, etc., or the granularity of the time axis, if the analysis target is sales. The interviews are repeated until all the necessary information is provided, and once all the information is available, the process proceeds to creating the detailed analysis specification document (S160).

[0058] If the analysis method is not time series forecasting but "other analysis methods (classification / regression / causal exploration, etc.)" (S150), proceed directly to creating the detailed analysis specification (S160).

[0059] Once the analysis method has been determined and the necessary information is gathered, an "analysis detailed specification" is created (S160) that summarizes the requirements necessary for the analysis, such as the analysis method and the scope and granularity of the data. At this stage, the analysis detailed specification organizes and describes the analysis objectives, project objectives, analysis method, and data to be used.

[0060] Next, it is determined whether the system is linked to a database (S170).

[0061] If the system is determined to be linked to a database (yes), based on the created detailed analysis specification, queries such as SQL are generated to extract the data necessary for the analysis from the database (S180). The generated queries are then used to actually extract the data, and after confirming the extracted data, it is formatted into a format suitable for analysis (S190). At this point, the process proceeds to the completion node, ending the processing flow in the consulting agent.

[0062] If it is determined that the system is not linked to a database (no), the process proceeds immediately to the completion node after creating the "Analysis Details Specification" (S160), and the processing flow in the consulting agent is terminated.

[0063] Specific examples of user-system interaction in a consulting agent are shown in Figures 7A and 7B.

[0064] First, the system asks the user about the analysis topic, that is, "what to analyze (analysis subject) and what objective you want to achieve (analysis objective)." The user might respond, for example, "I want to analyze factory power consumption to predict future power demand." If the information provided by the user is insufficient (e.g., only the analysis subject is entered, or the analysis objective is unclear), the system prompts the user to provide further information.

[0065] Once the analysis theme is determined, the system fills in a template for the detailed analysis specification to the extent possible according to the theme and presents it to the user, while also asking the user questions about any missing information or whether any corrections are needed to the information already entered. The information included in the detailed analysis specification template based on the user's responses includes, for example, the industry and scope of the analysis, the required analysis results, the available data, and how the prediction results are planned to be used, as shown in Figures 7A and 7B.

[0066] Once all fields in the detailed analysis specification template are filled in, the system presents it to the user as a summary of the analysis policy and requests their confirmation. Upon user approval, the content of the detailed analysis specification is finalized. The user can request the system to make any necessary corrections or modifications to the proposed detailed analysis specification if there are any deficiencies or issues.

[0067] [Data Understanding Agent] The processing flow in the data understanding agent is shown in Figures 8A and 8B. The data understanding agent receives the detailed analysis specifications from the consulting agent, performs preprocessing such as removing missing values ​​and estimating the scale of each variable, and prepares for the analysis before handing over the detailed analysis specifications to the analysis agent.

[0068] Starting from the starting node, the first step is to determine whether a CSV file needs to be uploaded in the "CSV Upload Necessity Determination" (S200). If True, the process proceeds to "CSV Upload Request" (S201). This occurs when processing by the consulting agent is skipped, when processing by the consulting agent is performed but there is no suitable data in the database, or when the system is not linked to a database. If False, the process immediately proceeds to "Generate Initial Dataset Information" (S207).

[0069] If it is determined that a CSV file needs to be uploaded, the system prompts the user to upload the CSV file in the "CSV Upload Request" (S201), and then validates the uploaded CSV file in the "CSV Content Validation" (S202). If the validation result is True, the system proceeds to "Confirmation of Analysis Purpose" (S203); otherwise, it returns to the "CSV Upload Request" and prompts the user to re-upload the CSV file (S201).

[0070] Next, in "Confirmation of Analysis Objectives" (S203), proceed to one of the following three options. (1) "Analysis-type questions" (S204) (2) "Request for specification of dependent variable" (S205) (3) "Generating initial data set information" (S207)

[0071] If the process proceeds to the "Analysis Type Question," the user is asked about the analysis type (analysis method) (S204). Various analysis types are possible, such as prediction, classification, regression, causal exploration, effectiveness verification, correlation analysis, and cluster analysis, but here we will explain using prediction, causal exploration, and effectiveness verification as examples. After the user's response, the process may eventually return to "Confirmation of Analysis Objective" (S203), and the flow may branch.

[0072] If the process proceeds to "Request for Specification of Dependent Variable," the user is asked to specify the dependent variable (target variable) to be used in the analysis (S205). Then, the process proceeds to "Verification of Dependent Variable Response" to verify whether the target variable is valid (S206). If the verification result is True, the process proceeds to "Generate Initial Dataset Information" (S207). If it is False, the process returns to "Confirmation of Analysis Objective" (S203) and the flow is repeated.

[0073] In "Generating Initial Dataset Information," initial dataset information, such as the number of unique values ​​in each column, is generated (S207). Then, the process proceeds to "Time Series Determination by Logic," where the logic determines whether the dataset contains columns representing time series (S208). If the result is True, it is determined to be a time series, and the process proceeds to "Initializing the Scale Determination State" of the sub-agent (scale identifier) ​​(S250). If it is False, the process proceeds to "Time Series Determination by LLM" to have the LLM further determine whether it is a time series (S209).

[0074] In "Time Series Determination by LLM" (S209), LLM is used to determine whether a column representing a time series is included. If the result is True, the process proceeds to "Initialization of Scale Determination State" (S250). Here, the result is True when it can be determined that time series data is included, or when it can be determined that time series data is not included.

[0075] In the "Time Series Determination by LLM" (S209), if the LLM determines that the data may be time-series data but it cannot be certain, it returns False, and the process proceeds to "Time Series Determination by Human," where the user is asked about the meaning of the data in the column that may represent a time series (S210). After the user's response, the process returns to "Time Series Determination by LLM" (S209) to make a final determination.

[0076] After the possibility of a time series is confirmed, the sub-agent (scale identifier) ​​shown in Figure 8B determines how to handle the "scale level" of the data (nominal scale, ordinal scale, interval scale, ratio scale, etc.). This is determined for each column of the dataset, and this sub-agent is looped through for each column.

[0077] In "Initialization of Scale Judgment State," the internal variables and flags necessary for the sub-agent are set up to their initial state (S250), and then the process proceeds to "Selection of Scale Judgment Items" (S251). In "Selection of Scale Judgment Items," the user selects what kind of verification to perform (S251), and from here the process branches into several steps, and after confirmation in each case, the process returns to "Selection of Scale Judgment Items" (S251), repeating this flow. (1) In the case of "Skip binary and time series", if a binary variable (binary) or time series index column has already been determined, that column is skipped (S252). (2) In "evaluation of order," we evaluate whether there is an order to the values ​​of the variables (S253). (3) In "determining whether it is an ordinal scale," we determine whether it can be definitively identified as an ordinal scale (S254). (4) In "evaluation of equal intervals," we evaluate whether the values ​​of the variables are equally spaced (S255).

[0078] If additional information is needed from the user to determine the scale, create a question in "Create a question for the user" (S256), ask the user the question (S257), and then proceed to "Select scale determination items" again (S251).

[0079] Once all scale estimation is complete (S258), exit the sub-agent (scale identifier) ​​and proceed to "dataset type determination" (S211).

[0080] In "Dataset Type Determination," the type of dataset (suitable for time series analysis, categorical data, quantitative data, mixed data, etc.) is determined based on information such as the scale of the data and whether it is a time series (S211). Depending on the determination result, the process proceeds to "Analysis Method Determination," where the type of algorithm to be used for analysis is further determined (S212).

[0081] In "Analysis Method Determination," the type of algorithm to be used for analysis is determined (AutoML for time series forecasting, AutoML for forecasting and classification, causal search for quantitative data, causal search for categorical / mixed data) (S212). If causal search for categorical / mixed data is selected, the result is True, and the process proceeds to "Question about the ideal value of the dependent variable" (S213). If any other method is selected, the result is False, and the process proceeds to "Distribution Determination" (S215).

[0082] In the "Asking about the ideal value of the dependent variable" step, the user is asked about the ideal value of the target variable (S213), and then the process proceeds to "Verifying the answer" (S214). This information is used when generating the explanatory text and reports of the causal exploration results.

[0083] In "Verify Response," the system checks whether the value of the target variable specified by the user is included in the dataset (S214). If the result is True, the system proceeds to "Determine Distribution" (S215); otherwise, it returns to "Ask about the Ideal Value of the Target Variable" and asks the user again (S213).

[0084] "Distribution determination" (S215) is the process of examining the distribution characteristics of the data. Here, outliers and the shape of the distribution are identified, and preparations are made for selecting the optimal analysis method. After this, the process moves on to "Handling missing values ​​and outliers" to decide how to handle missing values ​​and outliers (remove, impute, transform, etc.) (S216). This completes the preprocessing and analysis method selection flow.

[0085] Although not shown in the flowchart, the data understanding agent should include information regarding the analysis method and data used, as specified in the detailed analysis specification document.

[0086] Figure 9 shows a specific example of the interaction between the user and the system in a data understanding agent. For example, the system asks the user what the values ​​1, 2, and 3 of the variable name "Diagnostic Category" mean, and the user replies that they represent mild, moderate, and severe, respectively.

[0087] [Analysis Agent] Figure 10 shows the processing flow in the analysis agent.

[0088] Starting from the initial node, the process first branches into one of the following: "Prediction," "Causal Exploration," or "Effect Verification," based on the "Analysis Details Specification" received from the data understanding agent (S300). As mentioned above, "Prediction," "Causal Exploration," and "Effect Verification" are just examples of analysis types and are not the only ones that can be analyzed.

[0089] If "Prediction" is selected, the type of dataset (S: time series data with numerical dependent variable / numerical dependent variable / categorical dependent variable) is obtained from the detailed analysis specification (S310), the analysis algorithm (time series forecasting / regression / classification) is selected from the dataset type (S311), and the analysis is executed (S312).

[0090] If "Causal Exploration" is selected, the type of dataset (numerical only / category or mixed data) is obtained from the detailed analysis specifications (S320), the analysis algorithm (statistical causal exploration / score-constraint-based causal exploration) is selected from the dataset type (S321), and the analysis is executed (S322).

[0091] If "Effect Verification" is selected, the type of dataset and the results of user interviews are obtained from the detailed analysis specifications (S330), an analysis algorithm is selected based on the type of dataset and the results of user interviews (S331), and the analysis is executed (S332).

[0092] After performing analysis at all branches, the flowchart ends upon reaching a common completion node.

[0093] Although not shown in the flowchart, the analysis agent should include information regarding the analysis results and other items in the detailed analysis specification document.

[0094] Figure 11 shows a specific example of the user-system interaction in the analysis agent. When the system is ready to perform the analysis (S312, S322, S332) as shown in the flowchart in Figure 10, it may receive instructions to start the analysis, for example, by displaying a start button on the interface screen and prompting the user to press it.

[0095] [Result Interpretation Agent] Figure 12 shows the processing flow in the result interpretation agent.

[0096] Starting from the starting node, the analysis agent first branches into one of the following types of analysis: "prediction," "causal exploration," or "effectiveness verification" (S400). As mentioned above, "prediction," "causal exploration," and "effectiveness verification" are just examples of analysis types and are not the only ones that can be analyzed.

[0097] If "Prediction" is selected, the necessary information is obtained from the detailed analysis specification (S410), and a result interpretation statement for predictive analysis is generated (S411).

[0098] If "Causal Exploration" is selected, the necessary information is obtained from the detailed analysis specification (S420), and a result interpretation statement for causal exploration is generated (S421).

[0099] If "Effect Verification" is selected, the necessary information is obtained from the detailed analysis specification (S430), and an interpretation statement of the results for effect verification is generated (S431).

[0100] In all analysis procedures, after generating the result interpretation statement, the process proceeds to a step to check if there are any questions from the user (S440). If there are questions from the user, the process moves to a node to answer them (S441), and this process is repeated until there are no more questions.

[0101] Once there are no further questions from the user, or once all questions have been answered, the process reaches the completion node and exits the flowchart.

[0102] Although not shown in the flowchart, the results interpretation agent will include information regarding the analysis results and other items in the detailed analysis specification document.

[0103] Figure 13 shows a specific example of the user-system interaction in a result interpretation agent. The user asks questions about the result interpretation statement presented by the system, and the system answers them.

[0104] [Document Creation Agent] Figure 14 shows the processing flow in the document creation agent.

[0105] Starting from the starting node, the report generation process begins when the user clicks the "Next" button (S500) (S501). Note that this explanation uses the following chapter structure as an example, but the report format is not limited to this. Chapter 1: Executive Summary Chapter 2: Introduction Chapter 3: Data Used and Analytical Methods Chapter 4: Analysis results Chapter 5: Discussion of Analysis Results Chapter 6: Analyst Comments

[0106] First, based on the analysis results, Chapter 3 (Data Used and Analysis Methods) and Chapter 4 (Analysis Results) are prepared (S502, S503).

[0107] Next, based on the content of Chapter 4, relevant news articles are retrieved using an external API (S504), and Chapter 5 (Discussion of Analysis Results) is created by combining Chapter 4 and the retrieved news articles (S505).

[0108] Next, in addition to Chapters 3-5, we will generate Chapter 2 (Introduction), which will include a description of the background, objectives, methods, and results of the analysis, taking into consideration the current challenges, project objectives, data collection plan, and overview of the analytical methods (S506).

[0109] Next, chapters 2 through 5 are combined to create chapter 1 (executive summary), which serves as the overall introduction (S507). At this point, the content of the generated chapters 1 through 5 is displayed and ready for the user to review (S508). At this stage, the user can ask questions or request revisions regarding the content of chapters 1 through 5.

[0110] Once the content of chapters 1-5 is finalized after user confirmation, the generation process for chapter 6 (analyst comments) begins, using chapters 1-5 as input (S509).

[0111] First, news articles related to the content of chapters 1-5 are retrieved from an external API (S510).

[0112] Then, based on the content of Chapters 1-5 and the acquired articles, multiple analyst comments are created (S511). The analyst comments are written by multiple AI agents with different characters and expertise, each offering comments, insights, and policy suggestions on the analysis results from their respective perspectives.

[0113] Subsequently, the generated Chapter 6 (Analyst Comments) is displayed (S512), and the user is asked to review the content. If there are any questions or requests for corrections, these are addressed (S513). If there are no questions or corrections, or if all responses have been completed, the process proceeds to the next step.

[0114] Once the content of Chapter 6 (Analyst Comments) is finalized, the final report file is generated (S514), a download file is generated (S515), and provided to the user (S516). The report generation process is complete once the file download is finished.

[0115] Although not shown in the flowchart, the document creation agent will include information regarding the report items in the detailed analysis specification document.

[0116] Figures 15A and 15B show specific examples of user-system interaction in the document creation agent. When the user clicks the report generation button, the system creates chapters 1 through 5 of the report and presents them to the user. The user can then ask questions or request corrections regarding the content, and the system provides answers.

[0117] Thus, according to the data analysis system of the present invention, by utilizing multiple AI agents, users can perform a series of data analysis processes through dialogue with AI agents in a chat format without requiring specialized knowledge. As a result, processes that previously required specialized personnel such as data scientists to spend a considerable amount of time can now be performed in a short time even by non-experts.

[0118] Next, Embodiment 2 of the present invention is shown, in which the system of the present invention is applied to design review in the product development process. In design review in the product development process, multiple departments (planning, design, manufacturing, quality, etc.) are involved from planning to mass production, leading to discrepancies in the interpretation of required specifications, differences in impact assessment of design changes depending on the person in charge, making cost and delivery date predictions difficult, and past development know-how is dependent on individuals, resulting in insufficient utilization in future development. However, by using the system of the present invention, consistency in design review in the product development process can be ensured.

[0119] Figure 16 is a diagram showing the functional block of the computing device 110' according to Embodiment 2. As shown in Figure 16, the computing device 110' includes a requirements specification hearing agent 1200, a design policy selection agent 1300, a detailed design agent 1400, a design verification / prototyping agent 1500, a development result analysis agent 1600, and a decision-making document creation agent 1700. Note that the system configuration and the hardware configuration of the device are the same as in Embodiment 1, so a description is omitted.

[0120] The requirements specification hearing agent 1200 hears product development requirements from the user, automatically retrieves similar product data, market requirements data, technical constraint data, etc. from the design database, and defines development requirements (functional requirements, performance targets, cost constraints, etc.).

[0121] The design policy selection agent 1300 selects a design approach (modular design, reuse of existing parts, etc.) that is suitable for the requirements and determines the reasons for the selection (technical feasibility, cost-effectiveness, risk assessment, etc.).

[0122] The detailed design agent 1400 formulates specific design specifications (dimensions, materials, manufacturing process, etc.) from the design policy based on the results of design data analysis of similar products (strength calculations, thermal analysis, manufacturability evaluation, etc.).

[0123] The design verification and prototyping agent 1500 performs simulations (structural analysis, fluid analysis, manufacturability verification, etc.) based on the established design specifications and obtains verification results (performance evaluation, problems, improvement suggestions, etc.).

[0124] The Development Results Analysis Agent 1600 analyzes design and prototyping results in accordance with development goals (e.g., achievement of target performance, cost justification, manufacturing difficulty evaluation), and interprets success factors and challenges (e.g., design advantages, areas requiring improvement, mass production feasibility evaluation).

[0125] The decision-making document creation agent 1700 automatically generates product commercialization decision documents that organize development challenges, implementation details, and results.

[0126] In Embodiment 2, as in Embodiment 1, each AI agent can execute a consistent process by sequentially filling in the items in the "Detailed Analysis Specification" while engaging in repeated interactions with the user, and by sharing this information with other AI agents.

[0127] Although embodiments of the present invention have been described with specific examples above, the disclosed technology is not limited to the embodiments described above and can be broadly applied in various situations where multi-AI agents are utilized. It can be implemented in various other forms without departing from the gist of the disclosure. [Explanation of symbols]

[0128] 1…Data analysis system 10…Data analysis equipment 20…Administrator terminal device 30...User terminal device 40…Network 110, 110´...Arithmetic device 120...Storage device 130…Auxiliary storage device 140...Input device 150…Output device 160...Communication device 200... Consulting Agent 300...Data Understanding Agent 400…Analysis Agent 500…Result Interpretation Agent 600…Document creation agent 1200... Requirements specification hearing agent 1300…Design policy selection agent 1400…Detailed Design Agent 1500…Design Verification and Prototyping Agent 1600…Development results analysis agent 1700… Decision-making document creation agent

Claims

1. A data analysis system using a multi-AI agent in which multiple agents cooperate to perform a task, The aforementioned multiple agents include a consulting agent, a data understanding agent, an analysis agent, a results interpretation agent, and a document creation agent. A data analysis system in which each of the aforementioned multiple agents inputs information generated based on information obtained from the user in its own task into the item fields of the detailed analysis specification, and agents other than the consulting agent perform processing by referring to the information entered into the item fields of the detailed analysis specification by the other agents.

2. The data analysis system according to claim 1, wherein the multi-AI agent engages in a chat-like dialogue with the user.

3. The data analysis system according to claim 1 or 2, wherein each of the plurality of agents sequentially inputs information into the item fields of the detailed analysis specification, passes the detailed analysis specification to the next agent among the plurality of agents, and the agent performing the next processing performs processing by referring to the information entered into the item fields of the detailed analysis specification.

4. A data analysis method using a multi-AI agent in which multiple agents cooperate to perform a task, The aforementioned multiple agents include a consulting agent, a data understanding agent, an analysis agent, a results interpretation agent, and a document creation agent. A data analysis method in which each of the aforementioned multiple agents inputs information generated based on information obtained from the user in its own task into the item fields of the detailed analysis specification, and agents other than the consulting agent perform processing by referring to the information entered into the item fields of the detailed analysis specification by the other agents.

5. In an information processing device, Data analysis is performed using a multi-AI agent, in which multiple agents, including a consulting agent, a data understanding agent, an analysis agent, a results interpretation agent, and a document creation agent, work together to perform the task. A program that, in this case, causes each of the multiple agents to input information generated based on the information obtained from the user in their own task into the item fields of the detailed analysis specification, and causes agents other than the consulting agent to refer to the information entered into the item fields of the detailed analysis specification by the other agents and perform processing.

Images