Decision-making support device, decision-making support method, and program

Abstract

This decision-making support device comprises: an input unit that receives site information indicating the situation of a first site; a storage unit that records domain knowledge indicating a relationship among an issue at a training site, a first countermeasure implemented by a worker, and a site viewpoint evaluation index; a processing unit that inputs the situation of the first site and an issue at the first site to a large-scale language model that has referenced the domain knowledge to generate an output decision tree including a countermeasure candidate to be implemented at the first site and the site viewpoint evaluation index; and an output unit that outputs information including the output decision tree and a result of applying the countermeasure candidate to the first site.

Description

Decision support device, decision support method, and program

[0001] The present disclosure relates to a decision support device, a decision support method, and a program.

[0002] Conventionally, decision support devices that support decision-making for inventory planning in supply chains are known. As one of such decision support devices, it is provided with: an inventory arrangement plan generation unit that generates an inventory arrangement plan that achieves overall optimization of the supply chain; an allowable range constraint conversion unit that receives input of an allowable range for each of a plurality of participating enterprises constituting the supply chain with respect to evaluation indicators of the inventory arrangement plan, and generates constraint information; an inventory arrangement plan modification trial unit that attempts to modify the inventory arrangement plan using the constraint information; and a cause constraint identification unit that identifies the causal constraint information when a feasible solution cannot be obtained in the inventory arrangement plan modification trial unit; and an allowable range review request unit that requests the participating enterprise that has set the causal constraint information to re-input the allowable range (see Patent Document 1).

[0003] Japanese Unexamined Patent Publication No. 2023-136290

[0004] The decision support device of Patent Document 1 has room for improvement in ensuring the clarity of judgment criteria and preventing the occurrence of variations in decision-making by decision-makers.

[0005] The present disclosure provides a decision support device, a decision support method, and a program that can clarify the judgment criteria for decision-making and suppress variations in decision-making by decision-makers.

[0006] One aspect of this disclosure is a decision support device comprising: an input unit for inputting site information indicating the situation of a first site; a storage unit for recording domain knowledge showing the relationship between learning site challenges, first countermeasures implemented by workers, and evaluation indicators from a site perspective used to determine whether or not to implement the first countermeasures; a processing unit that generates an output decision tree including candidate countermeasures to be implemented at the first site and evaluation indicators from a site perspective used to determine whether or not to implement the candidate countermeasures by inputting the situation of the first site and the challenges at the first site into a large-scale language model that references the domain knowledge; and an output unit that outputs information including the output decision tree and the results of applying the candidate countermeasures to the first site.

[0007] One aspect of the present disclosure is a decision support method to be executed by at least one processor, comprising: inputting on-site information indicating the situation of a first site; recording domain knowledge showing the relationship between a learning on-site problem, a first countermeasure implemented by a worker, and an evaluation index from an on-site perspective used to determine whether or not to implement the first countermeasure; inputting the situation of the first site and the problem of the first site into a large-scale language model that references the domain knowledge, thereby generating an output decision tree that includes candidate countermeasures to be implemented at the first site and the evaluation index from an on-site perspective used to determine whether or not to implement the candidate countermeasures; and outputting information including the output decision tree and the result of applying the candidate countermeasures to the first site.

[0008] One aspect of this disclosure is a program for causing a computer to execute the decision support method described above.

[0009] According to this disclosure, the criteria for decision-making can be clarified, and variations in decision-making among decision-makers can be suppressed.

[0010] Figure 12 shows an example of a decision support system configuration. The diagram also shows an example of the operation of the decision support device, a flowchart, an example of the flow of countermeasures and considerations for a problem performed by an expert, an example of domain knowledge, an example of a decision tree based on domain knowledge, an example of a decision tree related to a technical method, an example of a decision tree related to management-perspective KPIs, an example of a prompt, an example of LLM output data using RAG, an example of a site-perspective decision tree that takes the target site into account, an example of a work plan screen, and another example of a site-perspective decision tree. The diagram shows a decision tree as a unit configuration of the site-perspective decision tree shown in Figure 12.

[0011] The embodiments will be described in detail below, with reference to the drawings as appropriate. However, unnecessary details may be omitted. For example, detailed explanations of already well-known matters or redundant explanations of substantially identical configurations may be omitted. This is to avoid the following explanation becoming unnecessarily verbose and to facilitate understanding by those skilled in the art. The accompanying drawings and the following explanation are provided to enable those skilled in the art to fully understand this disclosure and are not intended to limit the subject matter described in the claims.

[0012] (Background leading to the embodiment of this disclosure) In conventional decision support devices, the decision-making criteria for determining solutions to predetermined problems are subject to individual bias and are unclear. For example, while a skilled person familiar with a particular task may be able to perform the work more effectively, their knowledge may be limited to that of the skilled person, making it difficult for others to recognize the skilled person's decision-making criteria.

[0013] Furthermore, even if the judgment criteria of experts could be made visible, those criteria are based on on-the-job experience. These criteria may not necessarily align with those of management. In other words, even if management examines the judgment criteria of experts, it remains difficult to make judgments from a management perspective. Moreover, even among managers, the same criteria may not necessarily be used. Thus, there is variation in judgment criteria depending on the person making the decision.

[0014] The decision support device described in Patent Document 1 formulates an inventory plan that takes into account on-site optimization, but it does not provide the reasons why the inventory plan is the way it is. As a result, the decision support device described in Patent Document 1 requires workers to repeatedly try things out by trial and error, making it difficult to clarify the criteria for decision-making or to suppress variations in decision-making among decision-makers.

[0015] The following embodiments describe a decision support device, a decision support method, and a program that can clarify the criteria for decision-making and suppress variations in decision-making by decision-makers.

[0016] (First Embodiment) <Configuration of Decision Support Device> Figure 1 shows an example of the configuration of the decision support system 5 in the first embodiment. The decision support system 5 includes a decision support device 100 and a user terminal 200. The decision support device 100 and the user terminal 200 are connected to each other via, for example, a network 50.

[0017] The decision support device 100 includes a data storage unit 110, a knowledge storage processing unit 120, a processing unit 130, an input unit 140, an output unit 150, and a communication unit 160.

[0018] The data storage unit 110 stores on-site information related to the site. For example, the data storage unit 110 holds logistics warehouse information 111, manufacturing information 112, and sales information 113 as on-site information.

[0019] Logistics warehouse information 111 includes information about the logistics warehouse as a site. For example, logistics warehouse information 111 includes information about equipment in the logistics warehouse, information about workers working in the logistics warehouse, information about goods (products) handled in the logistics warehouse, or other information about the logistics warehouse. Logistics warehouse information 111 may also include past and current day receiving and shipping information, attendance information, etc.

[0020] Manufacturing information 112 includes information about manufacturing at the manufacturing site. For example, manufacturing information 112 includes information about equipment in the manufacturing site, information about workers working at the manufacturing site, information about goods handled at the manufacturing site, or other information related to manufacturing. Manufacturing information 112 may also include information about incoming and outgoing goods, attendance records, etc., for the past or for the day of work.

[0021] Sales information 113 includes information about sales at the store as a field. For example, sales information 113 includes information about equipment in the store, information about workers working in the store, information about goods (products) handled in the store, or other sales-related information. Sales information 113 may also include information about incoming and outgoing goods, attendance information, etc., from the past or for the day of work.

[0022] The data storage unit 110 may, for example, acquire data obtained from sensors or cameras installed on site as site information, or it may generate site information based on the acquired data.

[0023] The knowledge storage processing unit 120 comprises a knowledge unit 121, a countermeasure processing unit 122, a KPI unit 123, and a decision tree generation unit 124.

[0024] The Knowledge Unit 121 stores information related to various types of knowledge. The Knowledge Unit 121 executes various processes to generate domain knowledge and stores domain knowledge. Domain knowledge is knowledge held by people who are knowledgeable about a particular field or site, and in this embodiment, it is know-how and knowledge specific to the site. For example, domain knowledge is information that formalizes the knowledge of work held by workers (e.g., skilled workers). The Knowledge Unit 121 may, for example, hold the results of interviews with skilled workers at the site regarding work, or it may hold information from site issue management sheets and monthly reports.

[0025] Here, domain knowledge is defined as information that includes at least the relationship between (1) issues that may occur in the field, (2) countermeasures implemented to address those issues, and (3) evaluation indicators used to decide whether or not to implement those countermeasures. Technical methods, parameters, priorities, or evaluation indicators from a management perspective are elements to extend domain knowledge, but are not essential components.

[0026] The countermeasure processing unit 122 controls the system to derive countermeasures implemented by workers (e.g., skilled workers) in response to problems that occurred in the past at the worksite. The countermeasure processing unit 122 also controls the system to derive technical method information related to the technical methods for implementing each countermeasure. The countermeasure processing unit 122 stores the derived countermeasure information and technical method information. The countermeasure and technical method information here is based on past performance.

[0027] The KPI unit 123 controls the system to derive KPIs (key performance indicators) used to determine whether or not to implement countermeasures. The KPIs include multiple types, such as on-site perspective KPIs and management perspective KPIs. KPIs are one of the evaluation indicators used to evaluate various countermeasures. On-site perspective KPIs are indicators referenced by on-site workers, on-site managers, on-site supervisors, etc. On-site perspective KPIs are mainly referenced to evaluate whether work is being carried out efficiently. Management perspective KPIs are KPIs from a management perspective and are indicators referenced by managers and those involved in management. Management perspective KPIs are mainly referenced to evaluate whether the on-site financial situation is good or not. Therefore, the evaluation of countermeasures may not match between on-site perspective KPIs and management perspective KPIs. For example, measures such as introducing expensive equipment to the site may be highly effective from a site-level KPI perspective because it improves work efficiency, but it may be less effective from a management-level KPI perspective because it increases operating costs. The KPI unit 123 stores the derived KPI information. The KPIs here are KPIs based on past performance.

[0028] While operational KPIs are primarily used to evaluate the feasibility and efficiency of tasks, management-focused KPIs are used to evaluate costs, profitability, and alignment with business strategy. Although the same indicators may be used for both operational and management-focused KPIs, their evaluation objectives and perspectives differ.

[0029] The decision tree generation unit 124 generates a decision tree based on the above-mentioned past performance. The decision tree includes a field-perspective decision tree containing information on countermeasures from a field perspective for the problem. The decision tree also includes a management-perspective decision tree containing information on countermeasures from a management perspective for the problem.

[0030] The knowledge unit 121, the countermeasure processing unit 122, and the KPI unit 123 obtain response data by having the large-scale language model refer to predetermined information. The knowledge unit 121, the countermeasure processing unit 122, and the KPI unit 123 may obtain data such as the domain knowledge, countermeasures (candidate countermeasures), technical method information, and KPIs as response data. The large-scale language model includes, for example, RAG (Retrievable Augmented Generation) and LLM (Large Language Model).

[0031] In this embodiment, an example configuration for generating output decision trees using LLM or RAG has been provided, but the invention is not limited to this. For example, the generation of output decision trees may be achieved by rule-based inference methods, statistical models, machine learning models, or knowledge-based inference methods. In other words, the decision support device according to this disclosure does not require the use of LLM as its internal implementation, as long as it can structure and derive the relationship between judgment criteria and candidate countermeasures.

[0032] RAG is a technology that combines information retrieval and generation. RAG acquires relevant information from external data (for example, data provided by the Knowledge Unit 121, the Countermeasure Processing Unit 122, and the KPI Unit 123) and uses this relevant information to generate natural-sounding text using a generation AI such as LLM. RAG enhances the knowledge of the generation AI, enabling it to output more accurate response data.

[0033] The processing unit 130 includes a planning unit 131, a problem solving unit 132, an optimization unit 133, an effectiveness verification unit 134, an effectiveness verification unit 135, and a decision-making unit 136.

[0034] The planning unit 131 performs processing related to the formulation of a plan (for example, a plan for a specific task).

[0035] The problem-solving unit 132 performs processing related to problem-solving. The problem-solving unit 132 may be controlled to solve the problem via a problem-solving solver. For example, the problem-solving unit 132 may be controlled to generate a field-perspective decision tree corresponding to a predetermined site. The field-perspective decision tree includes candidate countermeasures for the problem and field-perspective KPIs for the candidate countermeasures.

[0036] The optimization unit 133 performs optimization-related processing. For example, the optimization unit 133 sends predetermined data to an optimizer to request it to optimize the resources at the work site and obtains the optimization results. Optimization is a method for obtaining the best possible result under given conditions, and it seeks a solution using numerical calculations or algorithms by maximizing or minimizing an objective function. Resource optimization includes, for example, the optimization of worker assignments, equipment used for work, equipment used for work, arrangement of items to be worked on, work sequence, work time, etc.

[0037] The effectiveness verification unit 134 generates a decision tree of management perspectives corresponding to a predetermined site. The decision tree of management perspectives includes candidate countermeasures for the problem and management perspective KPIs for the candidate countermeasures. The effectiveness verification unit 134 also determines the effectiveness of the candidate countermeasures for the problem, indicating how effective they are. The effectiveness verification unit 134 may determine the effectiveness of the candidate countermeasures after the site resources have been optimized by the optimization unit 133. The optimization may differ for each candidate countermeasure.

[0038] The effectiveness verification unit 135 determines whether the proposed countermeasures for the problem are effective (for example, whether they are feasible to implement). The effectiveness verification unit 135, for example, instructs the simulator to virtually implement the proposed countermeasures, that is, to perform a simulation, and obtains the simulation results (response data regarding effectiveness).

[0039] The decision-making unit 136 performs processing related to the decision made by the decision-maker.

[0040] The input unit 140 receives various types of data and information. Input by the input unit 140 may be by reading data without user intervention, or it may be by accepting user input when using the decision support device 100.

[0041] The output unit 150 outputs various types of data and information. The output from the output unit 150 may be displayed by the display unit, or it may be output in a format other than display (for example, audio output).

[0042] The communication unit 160 communicates various types of data and information. The communication unit 160 communicates with external devices (for example, user terminals 200). The communication unit 160 may communicate via the network 50 or without using the network 50. The communication unit 160 may communicate directly with the communication destination or via other communication devices.

[0043] The user terminal 200 is owned by a user. The user is, for example, a decision-maker, an employee, or a manager. The user terminal 200 may have a processor, memory, a communication device, an input (operation device), and a display device, just like a general terminal. The user terminal 200 may transmit data and information input to the user terminal 200 to the decision support device 100 and use it as input to the decision support device 100. The user terminal 200 may also receive and acquire data and information output from the decision support device 100 and output it (for example, display or audio output). In other words, the user terminal 200 can be used as an input / output terminal to assist the decision support device 100.

[0044] For example, the decision support apparatus 100 may be treated as a server apparatus, and the user terminal 200 may be treated as a client terminal. In this case, a user basically uses the user terminal 200. Alternatively, the decision support apparatus 100 may be treated as a PC (Personal Computer), and the user terminal 200 may be treated as a smartphone. In this case, the user may primarily use the decision support apparatus 100 and use the user terminal 200 in an auxiliary manner. When the decision support apparatus 100 is a server apparatus, the server apparatus may be configured as an on-premises server apparatus, or may be configured as a cloud type on a network. The processing performed by the decision support apparatus 100 may be performed by a single computer, or may be performed in a distributed manner by a plurality of computers.

[0045] It is to be noted that the functions of a large-scale language model (for example, an LLM), an optimizer, and a simulator are illustrated as being provided by an external server outside the decision support apparatus 100, but the present invention is not limited thereto. These functions may be provided inside the decision support apparatus 100.

[0046] Further, the decision support apparatus 100 includes, as a hardware configuration, for example, a processor, a memory, and a communication device. The decision support apparatus 100 may also include an input device (operation device) and an output device. The hardware configuration of the decision support apparatus 100 implements the above-described functions of the decision support apparatus 100.

[0047] The processor implements various functions by executing a program held in the memory. The processor comprehensively controls each unit of the decision support apparatus 100 and performs various processes.

[0048] The memory includes, for example, Random Access Memory (RAM) or Read Only Memory (ROM), and may include other memories and storages. The memory may be an external storage medium, and may be detachable from the decision support apparatus 100. The memory stores various data, information, programs, databases, and the like.

[0049] Communication devices transmit various types of data or information according to wired or wireless communication methods.

[0050] The input device may include various buttons, keys, a mouse, a keyboard, a touch panel, a microphone, or other input devices. The input device accepts input of various data or information.

[0051] Output devices include, for example, displays, speakers, and printers. Output devices output various types of data or information (e.g., display, audio output, printing).

[0052] Next, the operation of the decision support device 100 will be described.

[0053] Figure 2 is a flowchart showing an example of the operation of the decision support device 100.

[0054] The knowledge unit 121 collects domain knowledge (step S101). Specifically, the knowledge unit 121 collects and formalizes knowledge related to problem solving from experts familiar with the prescribed tasks and operations. Formalization refers to the process of converting tacit knowledge into a formalized form (making it explicit), including documenting and creating manuals. This makes it easier to share and transmit knowledge. Domain knowledge is obtained through formalization. Domain knowledge is, for example, information in which multiple pieces of information are related, and may be stored in the form of a table (see Figure 4) or a decision tree (see Figure 5).

[0055] The knowledge of experts is also called expert knowledge. Expert knowledge is experiential knowledge based on experience, difficult to put into words, and is primarily based on experience. In order to utilize expert knowledge, it must first be converted into tacit knowledge. Tacit knowledge is knowledge that can be written into text and can be formalized into knowledge that is easily transmitted to others.

[0056] Figure 3 shows an example of a workflow for countermeasures and considerations for a problem performed by an expert. For example, the workflow in Figure 3 allows for the acquisition of expert knowledge.

[0057] First, the expert assesses the situation at the site based on site inspections and reports and communications from relevant parties (e.g., the person in charge at the site) (Step S11). Based on their past experience, the expert identifies issues related to the assessed situation at the site (Step S12). Based on their past experience, the expert decides on countermeasures for the identified issues (Step S13). The expert or the person in charge at the site, as instructed by the expert, implements the decided countermeasures (Step S14). The expert confirms the results of the implemented countermeasures (Step S15). If the countermeasures were implemented by the person in charge at the site, the expert may confirm the results of the implemented countermeasures by receiving a report from the person in charge that includes the results of the countermeasures or by conducting a site inspection.

[0058] The KPI unit 123 estimates and formalizes the on-site perspective KPIs used by the skilled worker based on on-site information from the work performed by the skilled worker and the results of the countermeasures implemented at that site. On-site perspective KPIs are, for example, KPIs related to business efficiency and operations, such as process KPIs, customer KPIs, and employee KPIs. Some on-site perspective KPIs may be used consciously by the skilled worker, while others may be used unconsciously. Formalization may also be performed, for example, through interviews with the skilled worker or by the skilled worker themselves. In this case, the KPI unit 123 may input the results of the interviews with the skilled worker via the input unit 140 and formalize them. The KPI unit 123 may also acquire domain knowledge related to the skilled worker's knowledge using LLM or the like. In this case, on-site information and countermeasure information may be referenced by RAG to the LLM. The KPI unit 123 can reduce the effort required for formalizing domain knowledge by semi-automating the acquisition of domain knowledge using LLM, etc.

[0059] Figure 4 shows an example of Domain Knowledge DN1. Domain Knowledge DN1 includes, for example, information relating on-site issues, on-site perspective KPIs, countermeasures for the issues, and technical methods for implementing those countermeasures.

[0060] Figure 5 shows an example of a decision tree DT0 based on domain knowledge DN1. Since domain knowledge DN1 is based on the knowledge of experts, the decision tree DT0 is a decision tree in which knowledge from a field perspective is hierarchically structured.

[0061] In decision tree DT0, the top layer contains the problem. Below the problem, the field-perspective KPIs for that problem are described. Below the field-perspective KPIs, the countermeasures and technologies evaluated by the field-perspective KPIs are described. Decision tree DT0 also indicates that in order to improve field-perspective KPI 1, there are countermeasure 1 and technology 1 (technical method 1) to implement countermeasure 1, countermeasure 2 and technology 1 to implement countermeasure 2, and countermeasure 1 and technology 2 (technical method 2) and technology 3 (technical method 3) to implement countermeasure 1. Decision tree DT0 also indicates that there is a countermeasure 4 to improve field-perspective KPI 1, and that countermeasure 4 is a countermeasure of a different lineage from countermeasures 1 and 2. Decision tree DT0 may be generated for each problem.

[0062] Furthermore, decision tree DT0 (and the output decision tree described later) does not merely show classification results, but represents the structure of the judgment criteria that lead to decision-making. Specifically, the intermediate nodes of the decision tree represent evaluation indicators for evaluating countermeasures, and the leaf nodes represent feasible countermeasure candidates. In addition, the path from the root to the leaf nodes in the decision tree represents the decision process that shows the reason why a particular countermeasure candidate is selected.

[0063] For example, the Knowledge Unit 121 stores information about the knowledge of skilled workers in the field as information that can be referenced by the RAG to the LLM. Then, the Knowledge Unit 121 instructs the LLM, which has referenced the information stored by the RAG, to create explicit knowledge information, and acquires and stores domain knowledge DN1 and decision tree DT0 as response data from the LLM using the RAG. Here, the Knowledge Unit 121 extracts information using the LLM, which references a high-level abstraction database via the RAG.

[0064] Here, we will describe in detail the extraction of information using LLM by referencing a high-level abstraction database via RAG.

[0065] First, the knowledge unit 121 creates abstract data based on documents (for example, internal company documents) stored in the data storage unit 110, and stores this abstract data. For example, the knowledge unit 121 converts the term "secondary inspection," which is used internally and described in an internal company document, to the commonly used term "receiving inspection," and creates abstract data, which is data representing an abstracted version of that term. In this way, the knowledge unit 121 functions as a database with a high degree of abstraction.

[0066] Next, the knowledge unit 121 creates a prompt PT1 based on the field information stored in the data storage unit 110. In this case, the knowledge unit 121 expresses the prompt with the same level of abstraction as the abstracted data.

[0067] Next, the knowledge unit 121 takes the created prompt as input, sends the information of prompt PT1 to the LLM, and causes the LLM to refer to the abstracted data via RAG. The knowledge unit 121 obtains response data based on the prompt. The response data is the output data AS1 of the LLM using RAG. Based on the response data, the knowledge unit 121 creates a decision tree DT0 corresponding to the domain knowledge. Therefore, the decision tree DT0 is created using the abstracted data as terminology.

[0068] Figure 6 shows an example of prompt PT1. Figure 7 shows an example of LLM output data AS1 using RAG.

[0069] Furthermore, the knowledge unit 121 may acquire abstract data generated by a human via the input unit 140 for the storage of abstract data and the creation of prompt PT1, or it may be automated to automatically store abstract data and create prompt PT1 using a data catalog or prompt engineering. Even in the case of automation, some manual work may remain, such as having a human verify the contents of the stored abstract data and the created prompt PT1.

[0070] Furthermore, the process of extracting information using LLM, which references such a highly abstract database via RAG, is also applied in steps S102 and S103.

[0071] Returning to Figure 2, the countermeasure processing unit 122 generates and registers (i.e., stores) technical method information (step S102). Specifically, the countermeasure processing unit 122 stores the technical method for implementing the countermeasure and the parameters necessary to satisfy the constraints in a high-level abstraction database. The countermeasure processing unit 122 extracts parameters from the high-level abstraction database using RAG and LLM. The parameters are parameters related to the countermeasure (e.g., increasing personnel) (e.g., the number of people to increase). Therefore, the countermeasure processing unit 122 can determine the countermeasure concretely by extracting the parameters.

[0072] Here, we will elaborate on the accumulation of knowledge regarding technical methods.

[0073] The countermeasure processing unit 122 formalizes information regarding the technical method by, for example, the following method. Specifically, the knowledge unit 121 may register a decision tree DT3 that derives a technical method for realizing countermeasures against a problem and specific values ​​of parameters related to the countermeasures. The knowledge unit 121 may, for example, automatically generate the decision tree DT3 based on accumulated technical specifications. The technical specifications may, for example, be stored in the knowledge unit 121.

[0074] Figure 8 shows an example of a decision tree DT3 related to a technical method. Decision tree DT3 is an example of technical method information.

[0075] In the decision tree DT3, the technical method is placed at the top layer. Below the technical method, constraints (constraint conditions) related to the technical method are placed. In Figure 8, there are multiple constraints, and each constraint is shown as a separate branch. For example, constraint 1 is a constraint on personnel resources, and constraint 2 is a constraint on time. Below the constraints, information about variable parameters is placed. This information includes the content of the parameter and the content of the parameter changes. The parameters here include specific values ​​for countermeasures against the problem. For example, parameter A is a parameter indicating an increase in personnel, and the content of the change is to add one worker. For example, parameter B is a parameter indicating the amount of time delay, and the content of the change is to delay the deadline by two hours. Below the information about the parameters, the problem to be solved by changing the parameters is placed. This problem is, for example, a deadline delay.

[0076] In other words, the decision support device 100 can store technical method information for deriving specific countermeasures for a given problem, for example in the form of a decision tree DT3. Based on the technical method information, the decision support device 100 can determine, for example, whether a deadline delay will occur if the number of workers is increased by one or two, and derive specific countermeasures that can solve the problem by repeatedly changing the parameters.

[0077] Returning to Figure 2, the KPI unit 123 registers (stores) management perspective KPIs (step S103). Specifically, the KPI unit 123 registers management perspective information, including KPIs (management perspective KPIs) necessary for management decisions for each type of business, business model, or job (collectively referred to as business type, etc.) in the field, and the priority of the management perspective KPIs, into a highly abstract database. The KPI unit 123 extracts management perspective information from the database, for example, using RAG and LLM.

[0078] In steps S102 and S103, the data handled and the data generated differ, but in both cases, extraction is performed using LLM, with the above-mentioned high-level abstraction database referenced by RAG. The data generated in steps S102 and S103 may be generated in the form of a decision tree, for example, or in the form of association information where each data is simply linked.

[0079] Here, we will elaborate on the accumulation of knowledge regarding KPIs from a management perspective.

[0080] The Knowledge Unit 121 stores management perspective information that shows the relationship between measures instructed by on-site managers to the site in the past and the management perspective evaluation indicators used by the managers to decide whether or not to instruct the measures. In other words, the management perspective information is based on the managers' past performance. The KPI Unit 123 generates a decision tree DT2 that shows the relationship between the industry, business type or job type of the site, management perspective KPIs, and priority, based on the information collected at the time of past management decisions stored in the Knowledge Unit 121. The KPI Unit 123 may generate the decision tree DT2 using AI (Artificial Intelligence), i.e., RAG and LLM, or it may generate it manually. In the case of manual generation, the KPI Unit 123 may generate the decision tree DT2 based on input from the user via the input unit 140, or based on input from the user via the input device of the user terminal 200. The output unit 150 controls the display of the generated decision tree DT2.

[0081] Figure 9 shows an example of a decision tree DT2 related to management-perspective KPIs.

[0082] In the decision tree DT2, the industry, business type, or job type (also simply referred to as "industry, etc.") of the site is placed in the upper layers. The industry, etc. of the site may be placed not only in one layer but also hierarchically across multiple layers. For example, in Figure 9, the "Logistics Center" in the top layer is a higher-level concept of the industry, etc. of the site, while the "Fulfillment Center" and "Repair Center" in the lower layers of the top layer are lower-level concepts of the industry, etc. of the site. In other words, it shows that a logistics center has both a fulfillment center and a repair center.

[0083] In Figure 9, management-oriented KPIs are placed below the industry type of the site. In this case, multiple management-oriented KPIs may be shown for each priority level, or they may be shown hierarchically for each priority level. For example, in Figure 9, if the industry type of the site is a "fulfillment center," the management-oriented KPIs are "logistics costs" and "prioritizing long-term strategy," indicating that "logistics costs" is a higher priority management-oriented KPI than "prioritizing long-term strategy." Also, if the industry type of the site is a "repair center," the management-oriented KPIs are "CO2 emissions" and "logistics costs," indicating that "CO2 emissions" is a higher priority KPI than "logistics costs." In other words, the information on these management-oriented KPIs includes information on whether cost is the priority or environmental priority is the priority.

[0084] Returning to Figure 2, the data storage unit 110 acquires and stores on-site information about the site targeted for planning at predetermined timings. The output unit 150 controls the system to present the on-site information (e.g., display, audio output) (step S104). In this presentation control, for example, the output unit 150 may present the on-site information, or it may transmit the on-site information to the user terminal 200 via the communication unit 160, causing the user terminal 200 to present the on-site information. The same applies to the "presentation control" and "display control" described later. For example, the output unit 150 displays on-site information as shown in area A1 of Figure 11. The on-site information includes, for example, information such as the number of workers on the day, the volume of goods, and the destination of shipments.

[0085] The planning unit 131 formulates a plan based on on-site information. The planning unit 131 controls the system to perform a plan creation simulation, which virtually executes the formulated plan (step S105). The plan creation simulation is a simulation that determines whether the work can be carried out according to the plan based on on-site information. The plan creation simulation is performed by a simulator. The planning unit 131 sends on-site information to the simulator via the communication unit 160, for example, and obtains the results of the plan creation simulation from the simulator via the communication unit 160. This plan creation simulation is, for example, a discrete simulation.

[0086] Furthermore, the planning unit 131 may be controlled to perform a plan creation simulation with optimized on-site resources. The optimization of on-site resources is performed by an optimizer. The planning unit 131 sends on-site information to the optimizer via the communication unit 160, for example, and obtains the results of the on-site resource optimization from the optimizer via the communication unit 160. Alternatively, the planning unit 131 may be controlled to perform only the optimization of on-site resources without performing a plan creation simulation.

[0087] The problem-solving unit 132 refers to the results of the plan creation simulation or optimization and determines whether or not there are any problems in the formulated plan (step S106). If a problem exists, that is, if a problem is discovered (Yes in step S106), the process from step S107 onwards is performed. On the other hand, if no problem exists, that is, if no problem is discovered (No in step S106), the process shown in Figure 2 may be terminated without performing the process from step S107 onwards.

[0088] For example, the problem-solving unit 132 may determine that a problem exists if the results of the plan creation simulation indicate that some problem arises or that constraints are not met. For example, the problem-solving unit 132 determines that there is a problem in the plan if it determines that the deadline cannot be met if the work is carried out according to the plan. In other words, the problem-solving unit 132 may determine that the problem at this site corresponds to a parameter that does not meet predetermined constraints as a result of simulating the site information using the simulator.

[0089] The problem-solving unit 132 controls the generation of a site-centric decision tree DT1 that is appropriate for the target site, based on the site information and the information on the discovered issues (step S107). In this case, the problem-solving unit 132 may generate the decision tree DT1 by inputting the site information and the issues of the target site into a large-scale language model (e.g., LLM) that references the domain knowledge DN1 via RAG. The decision tree DT1 includes candidate countermeasures to be implemented at the target site and site-centric KPIs used to determine whether or not to implement the candidate countermeasures. The site-centric decision tree DT1 is a decision tree for extracting candidate countermeasures and is a decision tree that reflects the site-centric KPIs. The decision tree DT1 is generated for each issue.

[0090] Here, we will describe in detail the generation of the decision tree DT1 in step S107.

[0091] The problem-solving unit 132 acquires domain knowledge DN1 from the knowledge unit 121. Domain knowledge DN1 includes, for example, information based on past countermeasures taken by experts. The problem-solving unit 132 acquires on-site information from the data storage unit 110. On-site information includes, for example, information such as the type of business at the site, company overview, size, role of the relevant site, business content, processes, work procedures, management goals, roadmap, social responsibility, contracts, resources, etc. On-site information may include, for example, information by period, and may include, for example, on-site information for today.

[0092] The problem-solving unit 132 uses RAG to have the LLM refer to the domain knowledge DN1 (specifically, information regarding countermeasures based on the domain knowledge DN1) and the field information, and sends a predetermined prompt to the LLM. The problem-solving unit 132 obtains response data based on the prompt from the LLM using RAG. The prompt here includes instructions such as, "Based on the countermeasures based on the domain knowledge, generate candidate countermeasures for the field indicated by the field information in the form of a decision tree. Place field-perspective KPIs in the intermediate nodes of the decision tree, and place candidate countermeasures in the leaf nodes." This response data is, for example, the data of the decision tree DT1.

[0093] Figure 10 shows an example of decision tree DT1 from a field perspective. The format of decision tree DT1 is the same as that of decision tree DT0. In decision tree DT1, explanations of matters that are the same as those in decision tree DT0 are omitted or simplified.

[0094] As shown in Figure 10, in the on-site perspective decision tree DT1 corresponding to the target site, the issue includes deadline delays, and the on-site perspective KPIs include human productivity, shipping lot size, and shipping instruction delays. Furthermore, in decision tree DT1, measures to improve human productivity include extending work deadlines, extending departure deadlines, increasing personnel, adjusting workload, adjusting orders, introducing automation, reviewing processes, and reviewing work procedures. Workload adjustment may be, for example, leveling on a daily basis. Order adjustment may include, for example, changing the scheduled shipping date.

[0095] The output unit 150 controls the system to present the acquired decision tree DT1. The output unit 150 may also display information for each type of measure in different ways (e.g., different colors) if the decision tree DT1 includes both measures included in the domain knowledge DN1 and measures not included in the domain knowledge DN1. Here, measures not included in the domain knowledge DN1 are, for example, measures generated by the LLM based on general knowledge. Furthermore, measures may include those that take a short time to implement but only have a temporary effect (temporary measures), such as optimizing on-site personnel allocation, and those that take a long time to implement but have a permanent effect (permanent measures), such as improving personnel communication and training. In this case, the output unit 150 may display temporary and permanent measures in different ways (e.g., different colors). Additionally, measures may include those that involve optimizing on-site resources and those that do not. Measures that involve optimizing on-site resources include, for example, optimizing on-site personnel allocation, while measures that do not involve optimizing on-site resources include, for example, introducing equipment to automate on-site work (automation introduction), improving communication, and training. In this case, the output unit 150 may display measures that involve optimizing on-site resources and measures that do not involve optimizing on-site resources in different ways (different colors, different areas, etc.). Permanent measures are often measures that do not involve optimizing on-site resources, but this is not always the case. Thus, the output unit 150 may determine the display method of the result of applying the decision tree DT1 or the target site (see Figure 11) based on the nature of the candidate measures.

[0096] Furthermore, by visually displaying the output decision tree and the comparison results of the proposed countermeasures, the decision-maker can quickly grasp the judgment criteria and the reasons for selecting the countermeasures. This shortens the time required for decision-making, improves the reproducibility of the decision-making results, and suppresses variability in judgments among decision-makers.

[0097] The decision support device 100 can generate and present a decision tree DT1 that includes on-site perspective KPIs corresponding to the target site. Therefore, decision-makers can easily confirm what criteria were used to select potential countermeasures. Furthermore, on-site information may contain a mixture of clearly recognizable on-site perspective KPIs and on-site perspective KPIs that are not clearly recognizable. Even in this case, the decision support device 100 can derive a decision tree DT1 with clearly defined on-site perspective KPIs by having the LLM refer to the on-site information using RAG and supplementing the on-site information.

[0098] Returning to Figure 2, the optimization unit 133 extracts candidate countermeasures and technical method information for the problem discovered in step S106 based on the technical method information and the decision tree DT1 (step S108). In this case, the optimization unit 133 may use RAG and LLM to extract the above candidate countermeasures and technical method information. If there are multiple candidate countermeasures, for example, a candidate with a high priority or an arbitrary candidate countermeasure may be extracted.

[0099] The optimization unit 133 performs optimization by determining the parameters to be used for the countermeasure candidate in the technical method information (for example, the parameters shown in Figure 8) based on the countermeasure candidate and the technical method information (step S109). For example, the optimization unit 133 controls the system to optimize on-site resources by optimizing the parameters of the countermeasure candidate. In this case, the technical method information and the parameters of the countermeasure candidate are sent to the optimizer, and the parameter values ​​are obtained as response data.

[0100] The parameters here are setting values ​​that adjust the operation of the optimization (optimization algorithm), and may include, for example, the learning rate, the number of iterations, and initial values. The optimization unit 133 can achieve optimization by deriving appropriate values ​​as parameters, thereby improving the efficiency and accuracy of implementing candidate countermeasures.

[0101] The optimization unit 133 determines whether or not a predetermined constraint is met (step S109A). If the predetermined constraint is not met (No. in step S109A), the optimization unit 133 proceeds to step S108 and processes other candidate countermeasures.

[0102] If the predetermined constraints are met (Yes in step S109A), the effectiveness verification unit 134 verifies the effectiveness of the extracted countermeasure candidates (step S110). Specifically, the effectiveness verification unit 134 extracts management perspective KPIs corresponding to the target site based on management perspective information. Specifically, the effectiveness verification unit 134 uses RAG to have the LLM refer to the site information and management perspective information of the target site, sends a predetermined prompt to the LLM, and has it estimate the management perspective KPIs. Then, the effectiveness verification unit 134 calculates the estimated management perspective KPI values ​​based on the countermeasure candidates whose parameters have been determined.

[0103] Here, we will describe in detail the verification of effectiveness in step S110.

[0104] The effectiveness verification unit 134 estimates the management perspective KPIs corresponding to the target site based on the site information and management perspective information. For example, the effectiveness verification unit 134 refers to the decision tree DT2 shown in Figure 9, and if the target site is a repair center of a logistics center, it adopts "CO2 emissions," which is the management perspective KPI with the highest priority, as the management perspective KPI corresponding to the target site. The KPI unit 123 may also adopt multiple management perspective KPIs corresponding to the target site, in order of priority. The management perspective KPIs corresponding to the target site may be obtained in the form of a decision tree, as shown in Figure 9.

[0105] The effectiveness verification unit 134 calculates the value of the management perspective KPI for each candidate countermeasure included in the management perspective decision tree DT2 generated for the target site. In this case, the effectiveness verification unit 134 calculates the value of the management perspective KPI based on the parameter values ​​of the candidate countermeasure determined in step S109. In the example in Figure 9, the effectiveness verification unit 134 calculates the value of the management perspective KPI for each of the candidate countermeasures, such as extending the work deadline, extending the launch deadline, increasing personnel, adjusting workload, adjusting orders, introducing automation, reviewing processes, and reviewing work procedures.

[0106] Returning to Figure 2, the effectiveness verification unit 135 controls the system to perform an effectiveness verification simulation and determines whether the proposed countermeasure is effective or not (step S111). The effectiveness verification simulation is a simulation to determine whether the proposed countermeasure is feasible. The effectiveness verification simulation has stricter constraints than, for example, the optimization of the proposed countermeasure (optimization of on-site resources) described above. For example, various on-site constraints such as route information, worker leave information, and contract status are taken into account. Therefore, the effectiveness verification unit 135 determines whether the proposed countermeasure is feasible when the proposed countermeasure optimized in step S109 is executed.

[0107] If the proposed countermeasure is ineffective (No. in step S111), proceed to step S108 and continue processing other proposed countermeasures.

[0108] If the proposed countermeasure is effective (Yes in step S111), the output unit 150 is controlled to output (for example, display) decision information for decision-making based on the results of the effectiveness and practicality checks (step S112). For example, it is controlled to display the information of area A3 shown in Figure 11. This allows the decision-making unit 136 to provide decision-making information to the decision-maker.

[0109] The decision-maker confirms the displayed decision-making information. The decision-making unit 136 determines the countermeasure to be implemented from among one or more countermeasure candidates (step S113). In this case, the input unit 140 may receive input to select the countermeasure to be implemented from among one or more countermeasure candidates, and the decision-making unit 136 may determine the countermeasure to be implemented according to this selection information. Alternatively, the user terminal 200 may select the countermeasure to be implemented via an input device, transmit the selection information to the decision support device 100, and the decision-making unit 136 may acquire the selection information via the communication unit 160 and determine the countermeasure to be implemented according to this selection information.

[0110] According to this example of operation, the decision support device 100 can collect knowledge based on past performance in steps S101 to S103, and derive candidate countermeasures to solve problems that have actually occurred in the field in steps S104 to S111. The decision support device 100 can then provide information on the candidate countermeasures, including evaluation indicators from a field perspective and evaluation indicators from a management perspective, as judgment information for decision-making. Decision-makers can refer to the judgment information and make decisions while confirming the reasons why the candidate countermeasures were selected.

[0111] Next, I will explain the work plan screen.

[0112] Figure 11 shows an example of the work plan screen G1. The worker (for example, the decision-maker) works while looking at the work plan screen G1.

[0113] The work plan screen G1 includes areas A1, A2, and A3. Area A1 includes information on the work (task) content, simulation results based on on-site information, and simulation results after optimization of on-site information. Area A2 includes information on the on-site perspective decision tree DT1 corresponding to the target site. Area A3 includes information on optimization countermeasures (candidate countermeasures). Specifically, area A3 includes information on each countermeasure candidate included in decision tree DT1, the values ​​of each on-site perspective KPI when each countermeasure candidate is implemented, and the values ​​of the management perspective KPI when each countermeasure candidate is implemented.

[0114] In area A1, the simulation, which optimized on-site resources based on on-site information as of 18:00 the previous day, shows that zero deadline delays occurred (i.e., none occurred). In other words, the issue of deadline delays has been resolved through optimization. Therefore, there are no problems with the devised work plan, and the work plan will not be changed.

[0115] In area A1, the simulation, which optimized on-site resources based on on-site information as of 8:30 AM on the day, shows that 19 instances of deadline delay risk arise. This indicates that optimization alone cannot solve the problem. In this case, the work plan will be changed.

[0116] In response, the processing unit 130 searches for recovery plans (candidate countermeasures) to solve the problem based on the domain knowledge DN1, and generates a decision tree DT1 from a field perspective that is suitable for the target site. In area A2, the generated decision tree DT1 is shown.

[0117] The processing unit 130 then optimizes on-site resources for multiple candidate countermeasures and runs simulations. As a result, if there are zero instances of deadline delays in area A3, then that candidate countermeasure can solve the problem of deadline delays. The processing unit 130 calculates the value of a management-perspective KPI through simulations for each of the multiple candidate countermeasures and displays the information of the calculated value. A management-perspective KPI is, for example, VCPU (Variable Cost Per Unit). Decision-makers can make decisions by checking the VCPU value as a management-perspective KPI for each candidate countermeasure. Note that the candidate countermeasures in the optimization countermeasures table for area A3 may be those included in the candidate countermeasures in the decision tree DT1 for area A2, and there may be candidate countermeasures that are not included in the decision tree DT1. The optimization countermeasures table shows an example of the results of applying each candidate countermeasure to the target site.

[0118] The processing unit 130 displays information regarding automation implementation as a candidate for countermeasures that do not involve optimization of on-site resources. Since it is not possible to display optimized information in area A3 for countermeasures that do not involve optimization of on-site resources, the processing unit 130 may display information regarding automation implementation in an area other than area A3 on the work plan screen G1, such as in the output unit 150. The same applies when the candidate countermeasure is communication or training. In the above example, the processing unit 130 displays candidate countermeasures that do not involve optimization of on-site resources in an area other than area A3, but permanent countermeasures and candidate countermeasures that do not involve optimization of on-site resources often coincide. Therefore, the processing unit 130 may display temporary countermeasures in area A3 and permanent countermeasures in an area other than area A3.

[0119] Furthermore, in the work plan screen G1, it is not necessary for all of areas A1, A2, and A3 to be displayed; for example, area A1 may not be displayed. In other words, by displaying areas A2 and A3, the decision-maker can check the decision tree DT1 and information about each countermeasure candidate (optimization countermeasure table), and by checking the displayed information, the decision-maker can decide which countermeasure candidate to select. Also, by checking area A2 in the work plan screen G1, the decision-maker can understand what the issues are, what the judgment criteria are, and what countermeasures are available. Furthermore, by checking area A3 in the work plan screen G1, the decision-maker can check, for example, in a list which countermeasure to proceed with for optimization and what the values ​​of the management perspective KPIs are. This allows the decision-maker to selectively check only the necessary information, reducing the burden on decision-making. The only operation required by the decision-maker is, for example, to select the countermeasure to be implemented from one or more countermeasure candidates at the end. There may be other operations required by the decision-maker.

[0120] Thus, the decision support device 100 of this embodiment can formalize and visualize the judgment criteria of skilled workers regarding on-site work. Furthermore, since the formalized domain knowledge of the decision support device 100 includes on-site perspective KPIs, the judgment criteria from an on-site perspective can be clearly defined. Therefore, decision-makers can easily recognize what judgment criteria they are using to select countermeasures and recognize issues. In addition, the decision support device 100 can use management perspective KPIs to determine whether the derived countermeasure candidates are beneficial or not. Therefore, the judgment criteria from a management perspective can be objectively shown, and variability in decision-making can be reduced.

[0121] For example, in supply chains, optimization based on the tacit knowledge of experts tends to be highly individualized. The decision support device 100 can suppress the individualization of valuable knowledge possessed by experts by clarifying and formalizing such tacit knowledge. The decision support device 100 accumulates domain knowledge DN1, which includes general on-site perspective KPIs used by the clarified experts. Based on the domain knowledge DN1, the decision support device 100 derives candidate countermeasures and derives and presents management perspective KPIs that are tailored to the on-site business environment, thereby supporting more detailed decision-making.

[0122] (Variations of field-perspective decision trees) An example of a field-perspective decision tree is shown in Figures 5 and 10, but it is not limited to this. The knowledge unit 121 may acquire information on field-perspective decision trees having other configurations.

[0123] Figure 12 shows another example of a decision tree DT4 from a field perspective. In Figure 12, explanations of matters similar to those in Figure 5 or Figure 10 are omitted or simplified.

[0124] In the DT4 decision tree, the top layer contains "issue identification indicators." These indicators are used to quantitatively understand specific issues and problems occurring on-site and to evaluate the situation. For example, an issue identification indicator could be the number of surplus personnel. Below the issue identification indicators in the DT4 decision tree, "thresholds" are placed. These thresholds are conditions for determining whether or not an issue identification indicator should be considered an issue. For example, an issue identification indicator is considered an issue if it satisfies this condition (threshold). The threshold could be, for example, "<2," meaning the value of the issue identification indicator is less than 2. There may be multiple thresholds for an issue identification indicator. Below the thresholds in the DT4 decision tree, "issues" are placed. These issues are problems that need to be solved. For example, an issue could be a shortage of personnel. There may be multiple issues for a given threshold. Below the issues in the DT4 decision tree, "causes of the issues" and "solutions" are placed. For example, the causes and solutions for each issue are placed in the sub-layers. Here, the cause of the problem is the fundamental factor that causes the problem. For example, the cause of the problem might be the high level of skills required for the work. Here, the countermeasure is the specific action or measure to address the problem. For example, the countermeasure might be increasing personnel. In the decision tree DT4, the "countermeasure for the cause of the problem" is placed below the cause of the problem. Here, the countermeasure for the cause of the problem is the specific action or measure to address the cause that causes the problem. For example, the countermeasure for the cause of the problem might be reducing the number of tasks by dividing the process. Note that some layers in the decision tree DT4 may be omitted.

[0125] Figure 13 shows a decision tree DT5 as a unit configuration of the decision tree DT4 from the field perspective shown in Figure 12. Decision tree DT5 is composed of a single explicit knowledge tree structure. Multiple thresholds, issues, or other layer information can be provided in the unit configuration decision tree DT5, and decision trees DT4 of various shapes may be generated. Note that some layers in decision tree DT5 may be omitted.

[0126] Thus, the decision support device 100 may employ decision tree DT4 or decision tree DT5 as a decision tree representing the perspective of the field.

[0127] (Summary of Embodiments) Based on the above, this disclosure contains at least the following information. The components etc. in parentheses are examples of those corresponding to the embodiments described above, but are not limited thereto.

[0128] (Item 1) A decision support device (decision support device 100) comprising: an input unit (input unit 140) for inputting site information indicating the situation of a first site (target site); a storage unit (data storage unit 110, knowledge storage processing unit 120) for recording domain knowledge showing the relationship between learning site challenges, first countermeasures implemented by workers, and site-perspective evaluation indicators used to determine whether or not to implement the first countermeasures; a processing unit (processing unit 130) for generating an output decision tree (decision tree DT1) that includes candidate countermeasures to be implemented at the first site and site-perspective evaluation indicators (site-perspective KPIs) used to determine whether or not to implement the candidate countermeasures by inputting the situation of the first site and the challenges of the first site into a large-scale language model (LLM) that references the domain knowledge; and an output unit (output unit 150) for outputting information including the output decision tree (information in area A2) and the results of applying the candidate countermeasures to the first site (optimization countermeasure table) (information in area A3).

[0129] This allows the decision support system to formalize the knowledge possessed by experts, for example, as domain knowledge, and utilize it to derive potential countermeasures and their evaluation indicators for the first site. By outputting an output decision tree, the decision support system can provide decision-makers with specific countermeasures and their evaluation indicators tailored to the first site. Decision-makers can then make decisions after reviewing the provided information. Therefore, the decision support system can clarify the criteria for decision-making and suppress variations in decision-making by decision-makers.

[0130] (Item 2) The decision support device according to Item 1, wherein the input unit further inputs management perspective information showing the relationship between a second countermeasure instructed by the on-site manager to the site and a management perspective evaluation index (management perspective KPI) used by the manager to decide whether or not to instruct the second countermeasure; the processing unit causes the large-scale language model, which references the site information and the management perspective information, to estimate the management perspective evaluation index to be evaluated at the first site; and the output unit further outputs a result based on the estimated management perspective evaluation index as a result of applying the countermeasure candidate to the first site.

[0131] As a result, the decision support system can estimate countermeasures and their evaluations from a management perspective, which differ from countermeasures and their evaluations from the perspective of on-site personnel, etc., in accordance with the situation at the first site. The decision support system can provide decision-makers with specific countermeasure candidates and evaluation indicators from a management perspective that are appropriate to the first site. Decision-makers can make decisions after reviewing the provided information. Therefore, the decision support system can clarify the criteria for decision-making by taking into account not only the on-site perspective but also the management perspective, and can suppress variations in decision-making by decision-makers.

[0132] (Item 3) The decision support device according to Item 1 or 2, wherein the storage unit stores information on a technical method for implementing the first countermeasure, the processing unit causes the large-scale language model, which references the domain knowledge and the information on the technical method, to generate the output decision tree, and the output decision tree includes information on a technical method for implementing the candidate countermeasure.

[0133] This allows the decision support system to also provide information on the technical methods for implementing proposed countermeasures. Decision-makers can then make decisions while reviewing these proposed countermeasures and their technical methods.

[0134] (Item 4) The decision support device according to any one of Items 1 to 3, wherein the output unit determines the display mode of the output decision tree or the result of applying the countermeasure candidate to the first site, based on the properties of the countermeasure candidate.

[0135] This allows the decision support device to quickly verify the properties of the proposed countermeasures. For example, the output unit may assign different colors within the output decision tree according to the properties of the proposed countermeasures, or it may change the display position of the proposed countermeasures (for example, displaying them outside the margin of the proposed countermeasures table).

[0136] (Item 5) The processing unit generates a plurality of candidate countermeasures to be implemented at the first site, and is a decision support device according to any one of Items 1 to 4.

[0137] This allows the decision support system to provide information on multiple potential solutions. Decision-makers can then compare these options and decide on a solution for the primary on-site problem.

[0138] (Item 6) The decision support device according to any one of Items 1 to 5, wherein the processing unit controls the optimization of resources for implementing the proposed countermeasures, and derives the result of applying the proposed countermeasures to the first site based on the results of the optimization.

[0139] This allows the decision support system to determine the effectiveness of proposed countermeasures while maximizing their results during implementation, and to provide information on the effectiveness assessment results. Based on this information, decision-makers can then select the most effective countermeasures as their chosen countermeasures.

[0140] (Item 7) The decision support device according to any one of Items 1 to 6, wherein the processing unit controls the device to perform a first simulation (effectiveness verification simulation) in which the candidate countermeasures are virtually implemented, and derives the result of applying the candidate countermeasures to the first site based on the result of the first simulation.

[0141] This allows the decision support system to determine the effectiveness of proposed countermeasures based on their hypothetical implementation, and to provide information on the effectiveness assessment results. Based on this information, decision-makers can then select the most effective countermeasures as their chosen solutions.

[0142] (Item 8) The decision support device according to any one of Items 1 to 7, wherein the processing unit controls the device to perform a second simulation (plan creation simulation) in which the first countermeasures implemented by the worker are virtually implemented based on the on-site information, and determines the issues corresponding to parameters that do not satisfy predetermined constraints in the results of the second simulation as the issues at the first site.

[0143] This allows the decision support system to automatically extract issues that align with the situation at the first site, thereby improving the objectivity of issue extraction.

[0144] (Item 9) A decision support method to be executed by at least one processor, comprising: inputting on-site information indicating the situation of a first site; recording domain knowledge indicating the relationship between a learning site problem, a first countermeasure implemented by a worker, and an evaluation index from a site perspective used to determine whether or not to implement the first countermeasure; inputting the situation of the first site and the problem of the first site into a large-scale language model that references the domain knowledge, thereby generating an output decision tree that includes candidate countermeasures to be implemented at the first site and the evaluation index from a site perspective used to determine whether or not to implement the candidate countermeasures; and outputting information that includes the output decision tree and the result of applying the candidate countermeasures to the first site.

[0145] This means that the decision-making support method will have the same effect as item 1.

[0146] (Item 10) A program for causing a computer to execute the decision support method described in Item 9.

[0147] This allows the program to achieve the same effect as item 1.

[0148] Although various embodiments have been described above with reference to the drawings, it goes without saying that this disclosure is not limited to such examples. It is clear to those skilled in the art that various modifications or alterations can be conceived within the scope of the claims, and these will naturally also fall within the technical scope of this disclosure. Furthermore, the components of the above embodiments may be combined in any way without departing from the spirit of the invention.

[0149] Furthermore, the above embodiment may also apply to a program that realizes the functions of a decision support method, which is supplied to a computer (for example, a decision support device 100) via a network or various storage media, and which is read and executed by the processor of this computer, as well as a recording medium on which this program is stored.

[0150] This application is based on a Japanese patent application (Patent Application No. 2025-067599) filed on April 16, 2025, the contents of which are incorporated herein by reference.

[0151] This disclosure is useful for decision support devices, decision support methods, and programs that can clarify decision-making criteria and suppress variations in decision-making by decision-makers.

[0152] 50 Network 100 Decision support device 110 Data storage unit 111 Logistics warehouse information 112 Manufacturing information 113 Sales information 120 Knowledge storage processing unit 121 Knowledge unit 122 Countermeasure processing unit 123 KPI unit 124 Decision tree generation unit 130 Processing unit 131 Planning unit 132 Problem solving unit 133 Optimization unit 134 Effectiveness verification unit 135 Effectiveness verification unit 136 Decision-making unit 140 Input unit 150 Output unit 160 Communication unit 200 User terminal G1 Work plan screen AS1 Output data of LLM using RAG DT0, DT1, DT2, DT3, DT4, DT5 Decision tree PT1 Prompt

Claims

1. A decision support device comprising: an input unit for inputting site information indicating the situation at a first site; a storage unit for recording domain knowledge showing the relationship between learning site challenges, first countermeasures implemented by workers, and evaluation indicators from a site perspective used to determine whether or not to implement the first countermeasures; a processing unit that generates an output decision tree including candidate countermeasures to be implemented at the first site and evaluation indicators from a site perspective used to determine whether or not to implement the candidate countermeasures by inputting the situation at the first site and the challenges at the first site into a large-scale language model that references the domain knowledge; and an output unit that outputs information including the output decision tree and the results of applying the candidate countermeasures to the first site.

2. The decision support device according to claim 1, wherein the input unit further inputs management perspective information showing the relationship between a second countermeasure instructed by the on-site manager to the site and an evaluation index of the management perspective used by the manager to decide whether or not to instruct the second countermeasure; the processing unit causes the large-scale language model, which references the site information and the management perspective information, to estimate the evaluation index of the management perspective to be evaluated at the first site; and the output unit further outputs a result based on the estimated evaluation index of the management perspective as a result of applying the candidate countermeasure to the first site.

3. The decision support device according to claim 1 or 2, wherein the storage unit stores information on a technical method for implementing the first countermeasure, the processing unit causes the large-scale language model, which references the domain knowledge and the information on the technical method, to generate the output decision tree, and the output decision tree includes information on a technical method for implementing the candidate countermeasure.

4. The decision support device according to claim 1 or 2, wherein the output unit determines the display mode of the output decision tree or the result of applying the countermeasure candidate to the first site, based on the properties of the countermeasure candidate.

5. The decision support device according to claim 1 or 2, wherein the processing unit generates a plurality of candidate countermeasures to be implemented at the first site.

6. The decision support device according to claim 1 or 2, wherein the processing unit controls the optimization of resources for implementing the proposed countermeasures, and derives the result of applying the proposed countermeasures to the first site based on the results of the optimization.

7. The decision support device according to claim 1 or 2, wherein the processing unit is controlled to perform a first simulation in which the candidate countermeasures are virtually implemented, and the result of applying the candidate countermeasures to the first site is derived based on the result of the first simulation.

8. The decision support device according to claim 1 or 2, wherein the processing unit controls the device to perform a second simulation in which the first countermeasure implemented by the worker is virtually carried out based on the on-site information, and determines the issues corresponding to parameters that do not satisfy predetermined constraints in the results of the second simulation as the issues at the first on-site.

9. A decision support method to be executed by at least one processor, comprising: inputting on-site information indicating the situation of a first site; recording domain knowledge indicating the relationship between a learning on-site problem, a first countermeasure implemented by a worker, and an evaluation index from an on-site perspective used to determine whether or not to implement the first countermeasure; inputting the situation of the first site and the problem of the first site into a large-scale language model that references the domain knowledge, thereby generating an output decision tree that includes candidate countermeasures to be implemented at the first site and the evaluation index from an on-site perspective used to determine whether or not to implement the candidate countermeasures; and outputting information that includes the output decision tree and the result of applying the candidate countermeasures to the first site.

10. A program for causing a computer to execute the decision support method described in claim 9.

Images