In recent years, due to financial difficulties in maintenance, there has been a trend for the government to outsource infrastructure facilities to the private sector. In addition, coupled with the expansion of institutional investors and funds that can easily assume long-term and large-scale risks, infrastructure investment has been attracting attention as an asset class. Infrastructure investment is characterized by low correlation of returns with other asset classes and has an inflation hedge effect, and it is expanding in scale as an alternative asset, especially overseas. On the other hand, infrastructure facility projects are generally long-term and involve various uncertainties, making it difficult to accurately estimate the asset value. Therefore, asset value is largely determined qualitatively through discussions between contracting parties, and it is difficult to judge the rationality of the value from a third-party perspective. In this study, we propose an asset value evaluation method aimed at providing investment information by considering the "deterioration phenomenon" of infrastructure, which is a characteristic property of the asset class. Furthermore, we empirically analyze the effectiveness of the proposed method using data from actual infrastructure facilities.

A major challenge for large language models (LLMs) is adapting to domain-specific knowledge and context efficiently. In this study, we propose a domain adaptation method using knowledge editing for sentiment analysis, where the criteria for positive and negative classifications change over time and across contexts (e.g., before and after the COVID-19 pandemic). While previous studies have demonstrated the effectiveness of knowledge editing, its use in sentiment classification remains underexplored. We explore a novel training approach leveraging knowledge editing and evaluate its effectiveness in a Japanese sentiment analysis task. Specifically, we propose an enhanced training method applying Rank-One Model Editing (ROME) to the LLaMA model and assess its performance in a zero-shot setting. The results show our approach achieves higher accuracy in sentiment classification than conventional methods. This study provides the first empirical validation of knowledge editing in sentiment analysis and highlights its potential as an efficient domain adaptation technique.

As a countermeasure against cyber-attacks on IoT devices, anomaly detection in communication using flow data such as IPFIX is being conducted. Obtaining communication data such as IPFIX or PCAP in an operating system requires additional software, which can be difficult to implement due to the potential impact on services. Cloud services such as AWS provide flow log capabilities, such as VPC Flow Logs, that allow data collection with minimal impact to services. However, these logs only contain information on unidirectional packet counts and byte counts, which makes them more difficult to handle compared to flow data because session information is spread across multiple records. Our previous research has proposed methods to improve the accuracy of anomaly detection using VPC Flow Logs by appropriately merging records into session units and converting them into bidirectional data. However, this method does not consider the presence or absence of responses to requests, which may affect the detection accuracy. This study proposes to split the data considering the presence or absence of responses for anomaly detection.

In assembly manufacturing lines, work analysis is required to understand and improve productivity. This involves measuring work time and checking the accuracy of work procedures. Traditionally, these measurements have been done manually through visual observation, which requires significant effort and makes automation of measurement and analysis challenging. Recently, research on Temporal Action Segmentation, which segments work videos into basic action units, has been actively conducted, and many deep learning models have been proposed. However, these models require high-cost annotations with action labels for each frame, posing a challenge for practical implementation. In this study, we examined the performance of action segmentation in assembly processes using a CLIP-based model called Caseg, which can input action labels as linguistic information. By providing action labels extracted from assembly standards as prompts, we achieved a certain level of accuracy in inference. This suggests the potential to contribute to the practical implementation of action segmentation while reducing annotation costs, and we report these findings.

In the development of medical image AI, the burden of collecting training data and creating annotations remains a significant challenge. This study proposes a novel approach utilizing SAM2 (Segment Anything Model 2), developed by Meta, as a foundation model to achieve high-precision lesion detection in endoscopic images with minimal annotated data. The proposed method enhances SAM2's superior segmentation capabilities with two key extensions specifically designed for endoscopic video challenges: a bidirectional prediction mechanism and a weight decay system. Through evaluation experiments using cystoscopic videos, we demonstrated that our approach achieves comparable or superior performance to conventional deep learning methods and models trained on large-scale datasets, despite using only a small number of annotations. These findings present a new approach for efficient dataset construction in medical image AI development.

In recent years, zero-shot editing methods using pretrained diffusion models have gained traction in computer vision, with increasing interest in their application to audio editing. Traditional audio editing approaches leverage cross-attention maps from diffusion models but fail to adapt to the diverse range of prompts, reducing their reliability in practical scenarios. In this study, we propose a novel audio editing framework that operates at the latent space level using the attention mechanism of diffusion models. Our method refines editing through cross-attention manipulation while optimizing the similarity between editing instructions and intermediate edited audio, ensuring precise alignment. We evaluate our approach against traditional methods on multiple audio clips, demonstrating that our framework achieves high editing accuracy while preserving the original structure, outperforming traditional methods in maintaining audio consistency.

Since the release of ChatGPT, generative AI with a large language model (LLM) has rapidly gained adoption, showing the potential to revolutionize various fields. This study examines a novel rock physics modeling approach, focusing on seismic velocity estimation in rocks saturated with CO2 or H2, using multiple LLM-based algorithms. As subsurface storage of these gases becomes increasingly important for the energy transition and decarbonization, it is crucial to monitor changes in rock elastic properties, such as velocity and density, after gas injection. Although existing theories, empirical relations, and laboratory experiments can estimate these changes, there remains uncertainty in selecting and calibrating the most suitable models. This work investigates whether multiple generic LLMs, trained without domain-specific data, can accurately select models and parameters for four gas-saturated seismic velocity problems. By conducting blind tests with published experimental benchmarks, this study evaluates whether text-based inputs and outputs can yield accurate rock physics estimations, thus offering a new avenue for rapid, data-light modeling in geophysical applications. The results demonstrate that, although the accuracy of responses varies depending on the algorithm and the question, in some cases the models can provide accurate answers within a short time.

A large volume of free-text data in electronic health records (EHRs) describes treatment discontinuations, including those caused by adverse events. However, because this information is insufficiently structured in existing databases and thus difficult to extract, it remains underutilized despite its significant value. In this study, we combined automated labeling using Large Language Models (LLMs) with a small amount of manual annotation to efficiently classify treatment discontinuations due to adverse events. We integrated approximately 6,256 LLM-labeled records with 200 manually annotated samples, then fine-tuned JMedRoBERTa and T5. When evaluated on a 100-record test set, the T5 model demonstrated high precision (0.83) but was limited to a recall of 0.25. Missing adverse events is a critical concern in clinical practice, underscoring the need for more extensive training data. In the future, we plan to expand our approach to other discontinuation reasons (e.g., patient preferences or insufficient therapeutic effect) and strive for practical clinical implementation.

Creating nurse schedules is a complex task that must account for a wide range of considerations, including the number of night shifts, the required combination of skills, and hospital-specific constraints. Although the field of mathematical optimization has long studied this issue as a “nurse scheduling problem,” translating real-world scheduling rules into a mathematical optimization model requires specialized knowledge, limiting its practical dissemination. In this study, we investigate a method to automatically convert scheduling rules written in natural language into a mathematical optimization model.While research on automatically formulating mathematical optimization problems with large language models (LLMs) has progressed, many real-world complex problems still pose challenges, partly due to the vast problem space involved. Our work narrows this space by focusing on the conditions commonly assumed in nurse scheduling problems, and then attempts to convert these rules into a mathematical optimization problem using an LLM. Preliminary validation suggests that it may be possible to generate nurse schedules solely from natural language input. We aim for this research to contribute to more efficient schedule creation in healthcare settings.

When creating manuals, it is important to design a table of contents (TOC) that makes it easy for users to find the necessary information. However, constructing a TOC based on the users' perspective is not straightforward, requiring specialized knowledge. To solve this problem, this study proposes an automatic TOC generation method using large language models. Specifically, we design prompts that incorporate the rules and expertise used by manual creation professionals, thereby generating a TOC structure in which users can easily find information. We compared the table of contents structure generated by the proposed method with the existing table of contents structure, and confirmed that it is equivalent to the structure created by experts. Furthermore, through the experiments conducted in this paper, we discuss the method’s potential to automatically generate a generic TOC structure, as well as possible directions for further enhancement.

In recent years, municipalities have increasingly utilized disaster response verification reports to enhance their disaster preparedness. These reports document near-miss incidents encountered by administrative personnel, and the collection and classification of such cases are essential for improving disaster management measures. Currently, case extraction is performed manually, followed by classification into 12 categories based on the context. However, even when limited to flood-related verification reports since 2017, there are 85 reports spanning a total of 7,053 pages, making manual extraction and classification increasingly unfeasible. Thus, a tool to support the extraction and classification of near-miss cases is in high demand. Our previous research proposed a BERT-based method for extracting near-miss incident sentences with high accuracy. This study extends that approach by classifying extracted sentences into 12 categories using BERT. The model is trained with manually labeled data. Evaluation using Top-3 accuracy shows that while performance for low-data categories remains a challenge, accuracy for high-data categories reaches a practically usable level.

As Large Language Models become increasingly widespread, countermeasures against attack methods such as jailbreaking and prompt injection have become an urgent issue. Existing defense methods like Safeguard Models, including Llama Guard, have been found to perform inadequately against attacks in Japanese. In this research, we developed AILBREAK, an attack dataset collection application utilizing gamification to enhance LLM defense capabilities against Japanese-language prompt attacks. This application implements a mechanism that collects manually crafted attack prompts from users, featuring stages composed of various challenges based on safety categories, such as extracting passwords from enemies through battle-game elements. The design achieves both educational impact and data collection objectives. The collected dataset will be made publicly available for improving LLM defense functions and developing Japanese language-specific Safeguard Models. This paper reports on the application design, data collection methodology, and characteristics of the collected data.

Image scoring is a crucial task with many practical applications. To trust a model's judgement, understanding its rationale is essential. This paper proposes a novel training method for Vision Language Models (VLMs) to generate not only image scores but also corresponding justifications by natural languages. Leveraging only an image scoring dataset and an instruction-tuned VLM, our method enables self-training, employing the VLM's generated text without relying on external data or models. In addition, we introduce a simple method for creating a dataset designed to improve the consistency between predicted scores and their textual explanations. Through iterative training of the model with Direct Preference Optimization on two distinct datasets and merging them, we can improve both scoring accuracy and the coherence of generated explanations.

In the development of laboratory automation and laboratory RAG (Retrieval-Augmented Generation) systems, it is crucial to convert PDFs, such as laboratory equipment manuals, into formats that are easier for large language models (LLMs) to process. However, document layout analysis services based on deep learning often misinterpret human-optimized elements, such as UI screens and button images, treating them as tables or text. Additionally, binary analysis of PDFs is frequently inadequate due to the diverse ways PDFs are generated. To address these challenges, this study fine-tuned DocLayout-YOLO, pre-trained on DocSynth300K, using a dataset of 35,535 PDFs and annotations generated from HTML manuals for home appliances and electronic devices. The developed model achieves high accuracy in detecting UI and button images, even in complex cases. Furthermore, we propose a PDF analysis pipeline that integrates OCR and VLM to extract text, images, and structural information, converting the data into Markdown format. This work not only improves the efficiency of manual organization and reference but also provides a robust technological foundation for document processing in various fields.

Demand forecasting is an essential task in retail and manufacturing industries and has been the subject of numerous studies. Conventional popular time-series forecasting methods, such as the ARIMA model, require us to develop a forecasting model for each product.However, when products are frequently replaced and have short sales periods, we do not have enough data to build models individually.This study focuses on zero-shot time-series forecasting methods for demand forecasting with limited data. Zero-shot time-series forecasting is a framework for time-series prediction that does not require fine-tuning with specific time-series data to be predicted.To address the data shortage in practical situations, we propose a zero-shot demand forecasting model that considers exogenous variables. Our experiments with real data demonstrate that our proposed method achieved higher prediction accuracy than existing time-series forecasting methods, especially for products with short sales periods.

Japanese traditional publishers hadle a wide variety of content, including not only printed books, magazines, and comics, but also their electronic versions, digital media, and related merchandise. Traditionally, the production, sales, and distribution of these contents have relied heavily on past performance and experience of the people involved. The integration of AI has significant potential to realize more efficient and effective content creation, sales, and distribution processes. This combination of traditional expertise and cutting-edge technology could revolutionize the publishing industry, providing new opportunities to optimize operations and better meet the diverse needs of consumers. In this paper, we introduce 3 examples which have expanded work at Japanese traditional publishing house with AI: “demand forecasting for a first volume of a digital comic using data from a comic app”, “pv forecasting for an article at a web media from its title”, and “books recommendation from a web media's article contents”. Editors and sales staff use these AI-based systems for daily work and give us positive feedback.

In recent years, accurately identifying customers’ latent needs has become increasingly important. We propose a customer interview training system that utilizes a Large Language Models (LLM) as the interviewee. Our system dynamically updates prompts based on the progress of the conversation, ensuring that users can only extract latent needs if they ask appropriate follow-up questions. Experimental results demonstrated that, while the number of questions asked by participants increased by only 35% with the baseline method, it increased by an average of 148% with the proposed method. Furthermore, the number of questions required to elicit a single piece of information from the system was up to four times higher in the proposed method than in the baseline, indicating its effectiveness in promoting in-depth questioning.

This study proposes a data-driven strategy integrating generative AI and Jobs Theory. Jobs Theory provides a framework to understand the underlying motivations behind customer actions, while generative AI creates synthetic data based on these motivations. The hybrid model dynamically generates customer profiles and behavioral scenarios, enabling real-time and adaptive personalization strategies. For instance, a customer with the job of “relaxing after work” might receive offers for café discounts or streaming service benefits, while a customer aiming to “enhance family time” could be suggested family-friendly event invitations or leisure activity tickets. This model enhances targeting accuracy, improves customer engagement, and establishes a foundation for ethical and flexible data utilization.

Contrast effects refer to a psychological phenomenon where the relative value of a presented item changes based on comparisons with other items. This effect is widely utilized in the fields of marketing and pricing strategy. On the other hand, conventional recommendation systems mainly determine recommended items based on predicted user-item affinity, and thus fail to sufficiently consider factors that contribute to the relative attractiveness of an item, such as the order of presentation or the combination of items being compared. This paper proposes an evaluation function to realize contrast effect-aware recommendations in a two-sided recommendation system, and validates its effectiveness through the results of A/B testing in a real-world service.

Quantitatively evaluating how pricing strategies contribute to customer acquisition is crucial. In our company, product prices are structured as tables, which allow combinations of delivery days and set. However, since prices cannot be differentiated for customers at the price list level, it is impossible to estimate effect sizes through A/B testing. In this study, we grouped treatment and control groups for pricing strategies at the price list level and estimated acquisition effects using a causal inference approach.

In recent years, Marketing Mix Modeling (MMM) has gained attention as a method for analyzing aggregated marketing data in response to the growing trend of privacy protection and technological regulations. MMM is a type of market response analysis technique used to calculate the return on advertising investment and optimize budget allocation. This article introduces practical use cases in marketing consulting services and a machine learning platform designed to stably deliver complex analyses.

Outsourcing Marketing Mix Modeling (MMM) presents significant challenges, particularly the lack of transparency in algorithms. Common methods like path analysis and regression often fail to capture the complexity of time-series data, risking reduced accuracy in performance measurement. Additionally, non-disclosure of algorithmic details prevents marketers from explaining discrepancies or requesting improvements when results seem inconsistent. Further issues include long lead times, often taking months, and high costs. This presentation introduces an approach to address these challenges through in-house implementation. By ensuring algorithm transparency, enhancing analytical accuracy, and enabling faster feedback, we demonstrate practical solutions and key considerations with real-world examples.

LLM agents, which enhance LLMs with memory and behavioral modules for autonomous behavior, are gaining attention. AI alignment ensures AI adheres to human values, but agents should also autonomously adjust their values based on their environment. In society, values and behaviors are shaped by *habitus*—unconscious preferences formed through experience. Sociologist Bourdieu classified preferences into "naive" (practical) and "pure" (aesthetic). These preferences are closely tied to occupations. This study explores how agents' behaviors and experiences in a virtual environment shape their values. Six agents with different occupations lived in a virtual world, and interviews about their hobbies and preferences revealed that preferences evolved beyond occupational biases. The findings suggest that AI agents, like humans, can develop new values through experience, highlighting the importance of incorporating habitus into AI design to create more adaptive, human-like agents.

Recent AI advances achieve remarkable performance, yet modeling the human brain remains a crucial challenge. The Whole Brain Architecture Approach utilizes the Brain Reference Architecture (BRA) to integrate major brain regions but struggles with building large-scale brain-inspired systems. This paper introduces the Brain-morphic software Implementation & Testing System (BITS), a framework guiding developers from design to multi-faceted evaluation. BITS implements Hypothetical Component Diagram (HCD) components as Uniform Circuits (UC). Each UC can adopt either a Biophysically Constrained Model (BCM), which captures neuronal dynamics through spike-based architectures for higher biological fidelity, or a Biophysically Agnostic Model (BAM), relying on more abstract neural networks for computational efficiency. Leveraging the Function Realization Graph (FRG), BITS assesses task performance, alignment with neural activity, structural adherence to BRA, and lesion-based abnormal-condition simulations. Through iterative refinement, this approach balances biological plausibility with practical performance, thereby contributing to the advancement of robust, brain-inspired AGI.

When designing Brain Reference Architecture, it is essential to evaluate the dynamic properties of the pre-designed neural circuits before development and implementation, as this evaluation streamlines the implementation process. One significant challenge is the incomplete neuroscientific knowledge in certain brain regions. Therefore, we need a methodology that enables system-level behavioral analysis even with partially unknown neural mechanisms. In this study, we focus on motifs, the fundamental computational units of the brain, and propose a method for analyzing the dynamic behavior inherent in the system structure using qualitative reasoning. By implementing a qualitative neurodynamics model, we achieve both model simplification and motif componentization. The simulation of these combined components enables us to predict the system-level behavior of specific neural architectures. This method is particularly valuable as it enables behavioral analysis of neural systems that include circuits with incomplete neuroscientific knowledge.

In brain reference architecture (BRA)-driven development, architectures have been constructed for brain regions with relatively clear functions, such as the hippocampus and amygdala. The hippocampus plays a crucial role in spatial cognition and episodic memory, while the amygdala is essential for adaptive fear conditioning. In this paper, we survey the anatomical structures that connect the two brain regions by integrating the two BRAs of the hippocampus and the amygdala, which are adjacent brain regions. We also consider how the interaction between the two brain regions relates to the functions of spatial cognition and fear conditioning, providing insights into constructing a functional hypothesis.

Driven by recent advancements in sophisticated machine learning, integrated AI systems show increasing promise for real-world deployment. Nonetheless, explainability is essential for systems that function alongside human users, as it underpins trust and transparent decision processes. This study introduces a theoretical foundation anchored in ACT-R, serving as a blueprint for implementing cognitive models in applied settings. Focusing on phonological awareness in language development, it investigates how such models capture individual cognitive characteristics. A key premise is that participants may fail to detect errors in models mirroring their traits. Building on this assumption, we conducted an experiment wherein participants interacted with an immature phonological awareness model, analyzing erroneous responses to infer user-specific attributes and parameters. We also deployed an audio filter to emulate personal traits and examined its effect on model preferences. Findings corroborate our hypothesis and highlight cognitive modeling’s potential for interpretability and adaptation.

Digital twin refers to a digital replica of a physical object created in the cyber space of a computer, aiming to enable real-time monitoring, prediction, control, and optimization through the exchange of data and information between physical objects like machines and equipment and their corresponding digital twins. Recently, there has been increasing expectation for digital twins to address challenges in manufacturing processes, such as improving productivity, inheriting technologies, and optimizing processes. However, the lack of a simple implementation framework for constructing digital twins and data exchange has hindered practical application. In this study, we prototyped a digital twin framework for factory equipment models using the emulation software Emulate3D, and conducted anomaly detection by comparing the data from physical objects obtained from control devices with the behavior of the digital twin model. This presentation will summarize the results and discuss challenges related to full implementation and automation.

In recent years, the functional requirements for products have become increasingly sophisticated due to considerations for the global environment and demands for cost reduction. In the case of aircraft engines, the requirements for each component have also become more advanced, necessitating high assembly precision for subcomponent assembly. Therefore, assembling multiple subcomponents in appropriate combinations to meet quality requirements is a crucial practical challenge. In this study, we propose a method that constructs a model to predict the acceptance probability of the final product based on combinations of subcomponents using historical manufacturing data, and then explores optimal combinations based on these predictions. Through simulation experiments, we compared our proposed method with random matching to verify its effectiveness. This method is also expected to be applicable to optimization challenges across various industrial processes.

While the utilization of data in manufacturing environments continues to expand, the integration of domain knowledge remains crucial for practical implementation. We have systematically structured a machine learning framework that incorporates domain knowledge, which is called Informed Machine Learning (IML). Then, we focus on its practical applications in fault detection systems for semiconductor lithography equipment, which plays a vital role in semiconductor manufacturing processes. This equipment, responsible for transferring circuit patterns onto silicon wafers, consists of high-precision optical systems and multiple units requiring precise control. Notably, the conventional method of fault detection cannot be performed due to the limited occurrence of failures in our problem, where it is difficult to obtain adequate labeled data beforehand. To tackle such a problem, we propose a practical fault detection framework that integrates expert knowledge correlating failure modes with equipment states, combined with sensor-based condition monitoring.

Conventional shape generation techniques utilize Variational Autoencoders (VAE) to generate shapes based on statistical probabilities, which diverges from approaches where designers generate shapes based on physical quantities. Therefore, this study demonstrates that by combining Physics-Informed Neural Networks (PINN) with Conditional Variational Autoencoders (CVAE), it is possible to generate shapes based on the effects of thermal diffusion. This method is expected to promote shape generation that takes physical characteristics into account.

In thermal design for satellites, constructing an accurate thermal mathematical model is essential for precisely predicting the temperatures of onboard equipment. However, the current practice involves manually adjusting the parameters of the thermal mathematical model based on the results of thermal vacuum tests, which is time-consuming and costly. This study aims to enhance accuracy and efficiency by utilizing simulation-based inference (SBI) through deep learning to automate parameter adjustments. SBI using deep learning is a method that trains a deep generative model on parameters sampled from prior distributions and data generated via simulations, enabling parameter estimation that reproduces observed data without performing likelihood calculations. This approach accommodates complex and nonlinear models. Additionally, it provides posterior distributions for the estimated parameters, enabling a quantitative assessment of uncertainties. Numerical experiments using a small satellite model demonstrated the effectiveness of this approach.

In recent years, there has been much development of LLM-based AI agents. For example, in the sales process, AI agents are expected to improve business efficiency by selecting business partners and preparing documents based on the results of business negotiations. Meanwhile, the need for AI agents is increasing in the manufacturing industry. Data handled in the manufacturing industry is becoming increasingly large and complex, and in order to analyze data and obtain useful results, it is necessary not only to select the appropriate method for the problem, but also to perform coding work including preprocessing and parameter adjustment. However, users are not always familiar with analysis methods, and in some cases, it takes time to select a method or create an execution code. Therefore, this paper proposes a method to solve these issues by utilizing LLM. Specifically, the following two points will be discussed. (1) Algorithm selection utilizing information about the data to be analyzed (2) Prompt correction by feedback of execution resultsExperiments with artificial data confirm that fine tuning BERT allows the selection of an appropriate factor estimation algorithm with high accuracy. We also confirmed that correcting the prompts based on error statements during code execution can result in executable code.

This research aims to achieve a human-AI collaborative design and evaluation in system design activities. In previous study, we proposed a support tool for risk assessment using system and spatial design information provided by designers. However, it is challenging for designers to provide the design information with uniform abstraction level for necessary and sufficient assessment. In this paper, we experimentally study the feasibility of using generative AI for collaborative modification to uniformize the abstraction level of the design information in the previous study. The collaborative modification process is divided into four information processing functions. In each function, two large language models with different knowledge levels are examined to evaluate the impact of knowledge volume and retrieval methods. The evaluation using the use case and activity diagrams of the previous study suggests that knowledge retrieval methods, such as prompt engineering, are more essential than knowledge volume in achieving collaborative modification.

The objective of this research is to propose an ontology of information which is used to improve aircraft maintenance planning including predictive maintenance based on the MBSE (model-based systems engineering) approach. The proposed ontology captures the concept and the relationships that exist across multiple stages of the life cycle of civil aircraft maintenance. Recently, a data-driven aircraft health management method has been introduced to improve the efficiency of maintenance planning. However, it takes several years to a decade to obtain sufficient maintenance result outcome data after the aircraft enters service. By using the proposed ontology, valuable information can be extracted even when the amount of data is relatively small. In this paper, examples of causal relationships of problems extracted from the described model based on the proposed ontology are shown. This result is expected to be used for maintenance planning optimization.

We are developing an interference exposure system. This system can expose high fine periodic patterns without shading mask. To improve the operational stability and maintainability of the system, we investigate the enhancement of stage control accuracy and the estimation of anomaly causes by using models. In the investigation, the accuracy of the model is important. In this paper, we report on the improvement of the accuracy of model with the symbolic regression. A genetic algorithm was used to identify a function for the difference between stage and model. We were able to clarify the forces acting on the stage without the motors by interpretation of the obtained function. We thought that the difference between the model and the actual machine contains useful information for know-how and system improvement. We expect that the proposal in this study will provide an opportunity to obtain such knowledge.