We present Evidence-based Semantics as a simple and fundamental but novel problem in AI that has many application areas: to explain the meaning of observed events via collecting useful evidence. Evidence here is a piece of new information from the open real world, that is useful for explaining the meaning of the observed event, and the knowledge for entailing the observed events. This information may not be of the generality that is the goal of learning, but may be a piece of knowledge or an assertion in a one-time message. In this presentation, we propose two approaches to EBS and a couple of applied toy examples of these approaches for actions and utterances in real life.

Since the skills possessed by engineers do not always match the skills required for a job, visualization of the demand and supply of skills is significant for both macro resource allocation and career advancement of engineers. In this study, we construct a signed network that integrates the demand and supply of skills, and use network analysis techniques to clarify the multi-scale relationships among skills. We classify and visualize skills by extracting skill sets that focus on the largest clique in the signed network, applying a community extraction method hierarchically. We interpret the results and findings with considerable suggestions. Finally, we compare signed skill networks by industry to identify field-specific skill sets and provide insights for crossdisciplinary human resource development.

There is a great need for extracting the sequential patterns from data that represent the sequence of human decision making, such as purchase histories and game action histories, as it can lead to measures for improving various services. Decision trees, which are representative identification models in the machine learning field, are highly interpretable models that allow us to find the basis of decisions from the splitting node features of the learned models. On the other hand, decision trees do not retain time-sequential order relations in their learning, and it is difficult to extract a sequential pattern from them. On the other hand, models that repeat recursive input such as RNN can be trained while preserving the order relation of series data, but the interpretability of the trained model is low. Therefore, in this study, we propose a model that preserves the order relation of series data by adding a series constraint to the search range of splitting node features when training decision trees. Since this model preserves the series order of the features used in the branching process, it is possible to extract interpretable sequential patterns and important branches of decision making in the series from the trained decision tree. In this paper, we conduct experiments on real data and show the effectiveness of the model based on the extracted patterns.

This study proposes a method to identify the difference of the package technology that two semiconductor companies hold by using their published patents. The authors use BERT to generate the document vectors of the patents and obtain the 2-dimensional feature map by using t-SNE dimensionality reduction. In the map, we identify clusters with different plot densities depending on the two companies. Word2vec analysis determines how often what technological terms appear in each cluster, and engineers specify the difference. This combination of NLP analysis reveals: one company has many patents on light related devices, while the other does so on magnetic sensors. Engineers has finally confirmed the validity of this machine-only analysis results.

The semantics of natural language based on dependent type theory has provided a method for recognizing textual entailment, and type checking algorithms underlying such recognizing textual entailment systems are formulated for systems consisting of symbolic reasoning. Recently, systems with embedded neural networks have been proposed for natural language semantics based on dependent type theory. Neural DTS, which is derived from Dependent Type Semantics (DTS), is one such system.In this study, by defining a type checking algorithm for Neural DTS, we formulate a neural inference procedure for the propositions composed from atomic relational propositions and their negations by conjunction and disjunction. First, a classifier is trained on a dataset extracted from WordNet. Next, we embed the trained classifier in a type checking algorithm, replacing binary relation symbols with it. We then show that prediction concerning the propositions mentioned above can be made by means of type checking for these propositions.

Electronic texts in the medical field are used for research in natural language processing, including the study of Recognizing Textual Entailment in clinical texts using the compositional semantics system, ccg2lambda. One problem with existing systems is that they are unable to correctly determine implication relationships when input sentences contain medical domain-specific paraphrases, such as names of diseases. To solve this problem, a method is proposed that uses a named entity extractor for disease names and a disease dictionary to identify candidate paraphrase expressions lacking in theorem proving, and completes equivalence relations of disease names as axioms to the theorem prover. In this study, for the aforementioned method, we extend the module for deriving axioms. We also construct an inference test set that requires axiom injection of disease names and evaluate the inference system using the extended module.

Several studies have been proposed to analyze the factors that cause stock price fluctuations. It's known that utilizing textual data is effective in building models. In order to utilize text data in this method, it's necessary to quantify sentences using bag-of-words, etc., and at that time, concrete information is abstracted. Therefore, when it comes to investigating what specific factors caused the stock price fluctuation at a certain point in time, it's inappropriate to do so. In this case, it's desirable to construct a model using data other than textual data first, and then clarify the textual information to be focused on from the model. In this study, I first build a time-varying coefficient model using numerical data. Then, by focusing on the residuals of the model, we found the relationship between stock price fluctuations and specific factors expressed in text data at surrounding points in time.

In this study,we propose a model that appropriately predicts the probability that a horse will win and a method for betting on horses with the highest possible return based on the horse's odds information.Specifically, since horse racing data is a mixture of time series data and attribute data,we use a partially recurrent neural network,which can appropriately predict the odds of a horse winning,to predict the odds of a horse winning.Furthermore,we propose a method to search for the optimal betting strategy while minimizing the loss by not only predicting the winning rate but also betting based on the expected value of the odds.Finally,we simulate two types of betting,the conventional method and the proposed method,using actual horse racing data,where the winning probability is the probability of finishing in the top three positions.

In recent years, many images and texts related to fashion coordination have been posted on social media services. It has become common for users to select their outfits based on other users' posts. Social media services store a large amount of user, image, linguistic information, etc., that are expected to be an advantage in business by the appropriate utilization. In this study, we propose a recommendation method for fashion coordination posts based on both image and linguistic similarity. Specifically, at first, Image2StyleGAN is learned to measure the image similarity, and Doc2Vec is learned to measure the linguistic information similarity. Secondly, we obtain the two types of similarity rankings by the above methods and calculate their weighted sum with an adjustable parameter. This allows us to customize which information is more important for each user. Finally, we present how to utilize the proposed method using real-world fashion coordination service application data.

In recent years, more and more used smartphones have been bought and sold through online sales services in the used smartphone market, and it is desirable to utilize the large amount of transaction data accumulated in conjunction with these transactions when listing and purchasing used smartphones. Used equipment buyers can use this data to analyze market price trends and the factors that determine those prices, which can lead to optimal purchase strategies and pricing. However, used smartphones are handled by various sales services. For such a target problem, it would be possible to understand which factors contribute to the selling price if a prediction model could be constructed to explain the selling price based on various characteristics. In this study, we analyze price determinants using a model that incorporates the gradient boosting method, which is a model with high accuracy and interpretability, with the help of explainable AI. In this analysis, it is undesirable to apply a single pricing factor analysis model that could not consider differences in sales services, which has been the focus of previous analyses of the used equipment market. Therefore, we proposes an analytical model that employs the technique of explainable AI for the different price determinants among sales services. The proposed model is applied to analyze actual sales data of used smartphones and discuss the results.

While there is a well-developed model of accounting information systems (e.g., REA), as the business risk approach adopted in financial auditing suggests, auditors have to face the fraud occurring outside the system, especially the ones that management involves. In this research, we have developed models of actions in two fields: accounting of impairment of fixed assets and audit of accounting estimates. Accounting of impairment of fixed assets was selected because impairments are closely related to management decisions. And we chose the auditing standard on accounting estimates because the most challenging part of the audit of impairments is examining the fixed assets' future cash flows. These models are constructed as action decomposition trees: the goal action is accompanied by the method or methods to achieve the action. And the method consists of the component actions that ultimately achieve the goal action. We constructed the models, by embedding the standard provisions into action and method nodes. This research shows that action decomposition trees can express these standards with some complemental description rules and may also clarify the relationships between particular requirements. Moreover, with these models, auditors might effectively examine the accounting estimates.

We propose a method for estimating the condition of farmed shrimp in Recycling Aquaculture System (RAS) tank using CPS (Cyber Physical System). In this proposal, the amount of dissolved oxygen (DO, Dissolved Oxygen) is measured by an IoT devices installed in the tank, and the state of farmed shrimp in the tank is estimated from the change in DO. In order to confirm the condition of cultured shrimp, existing methods often estimate from images, but it is difficult to apply in a turbid water tanks. On the other hand, the proposed method uses the measured values of DO, so it can be applied even in turbid water tanks. We constructed this proposed method as a CPS and confirmed that the shrimp activity can be estimated.

The challenge of applying AI to manufacturing has been tried for many years. The application of AI to the engineering chain is also underway and it is still far from complete. This paper describes the latest technology and examples of applications of AI to design and development in the manufacturing engineering chain. The activities for deeper and wider application will be presented. The use of data from thermal flow analysis and experimental results to training would be simplified evaluation of new designs, and examples of various applications using AI based ROM (reduced order model), AI surrogate models, and deep learning, as well as applications to vibration of large structures would be presented. Mechanisms for accumulating causal analysis and AI training data and future efforts will also be discussed.

The purpose of this research is video anomaly detection for industrial devices that perform repetitive motions in multiple patterns. Deep learning technology has been developing rapidly in recent years and has attracted much attention for industrial applications. Unsupervised learning, in which only normal conditions are learned, is effective because it is easy to use even for objects for which it is difficult to define abnormal conditions in advance, and is beginning to be widely used in image inspection and other applications. On the other hand, there are timing anomalies (anomalies that can be detected only by considering the regularity of temporal changes in motion) that are difficult to detect using only still images, and unsupervised learning is required to detect such anomalies. When detecting video abnormality, it is relatively easy to detect timing abnormality if the device always repeats the same action. This is the reason why we have developed this method. In this presentation, we will discuss a highly accurate timing error detection method for industrial equipment that performs repetitive operations with multiple patterns, and report the results of a method that combines AE and LSTM and a method that uses PredNet to obtain high accuracy.

In recent years, there has been a growing demand for analysis of worker behavior from the viewpoint of labor shortage and improved work efficiency in factories such as automobile assembly. Behavioral analysis of assembly operations makes it possible to automate the measurement of the time required for each task and to confirm that operators are following the same procedures as in the manual. Due to this growing demand, temporal action segmentation using Deep Neural Networks (DNNs) has been widely studied as a new behavior analysis technology. On the other hand, standard benchmark datasets for temporal action segmentation often have a person in action or an object accompanying the action occupying a large area within the viewing angle of the video. On the contrary, videos of automobile assembly operations show a moving automobile that is larger than a worker, which may interfere with the analysis of the workers’ behavior. Therefore, this study applies several existing temporal action segmentation methods to this problem and verifies their effectiveness. Experimental results suggested the possibility of automating behavior analysis in automobile assembly operations.

In recent years, as end products have become more diverse, welding must also be performed under a variety of conditions. Arc welding machines offer the best waveform control in the user's environment to achieve highly efficient and high-quality arc welding. As a result, slight environmental changes may affect the welding results, necessitating readjustment of waveform control parameters. However, it is difficult for users to adjust the waveform control parameters because it requires knowledge of how parameters affect the welding result and empirical prediction. Therefore, this paper proposes a welding waveform control parameter generation model in which the desired welding result can be obtained by entering the desired score. We constructed a model which generates the waveform control parameters that produce the desired welding results by training LightGBM.

In manufacturing processes, various types of soft-sensors have been widely used for predicting important variables, including quality variables, that is difficult to be measured with hardware sensors. Various methods combining transfer learning and adaptive modeling have been proposed for quick soft-sensor adaptation after process maintenance. In conventional methods, it is necessary to switch soft-sensors using transfer learning to soft-sensors without using transfer learning to prevent negative transfer, which may cause performance deterioration of soft-sensors. However, it is difficult to appropriately determine the timing of soft-sensor switching. In this study, we propose a new transfer learning-based adaptive soft-sensor that can avoid the problem of soft-sensor switching. In the proposed method, soft-sensors are constructed using samples measured within the last fixing period as the target domain, and updating them sequentially when transfer learning is adopted. The usefulness of the proposed method was illustrated through its application to real data on the distillation process of a fluorinated telomer intermediate. It was confirmed that the proposed method improved the RMSE by an average of 17% compared with the conventional method.

Bid price optimization in Internet advertising is a very difficult task due to its high uncertainty. In this paper, we propose a bid price optimization algorithm focused on keyword-level bidding for pay-per-click sponsored search ads. The algorithm first predicts the performance of keywords as a distribution by modeling the relationship between ad metrics through a Bayesian network and performing Bayesian inference, and then outputs the bid price using a Bandit algorithm and online optimization. This approach enables online optimization that consideres uncertainty from the limited information available to advertisers. We conducted simulations on real data and confirmed the effectiveness of the proposed method on both open source data and data provided by an Internet ad-serving company.

In this paper, we present a method for estimating users' interests and preferences using Twitter followers. In recent years, Twitter, a microblogging service, has been attracting attention and is being used in various business fields. In Twitter, users' interests and preferences are considered to appear not only in their own posts but also in those of the users they follow. Among them, influencers post about specific topics with high frequency, so it can be said that the influencers that users follow are an important factor in estimating users' tastes and preferences. In this method, we focus on the Tweets of influencers followed by the followers of the target brand's Twitter account. After conducting a topic analysis of the influencers' Tweets, clustering them enables us to estimate the interests and preferences of the target brand's followers. By implementing this method, brands can help provide a better purchasing experience for their users.

For service providers, the users who have used the service in the past but have already left are important because they are still expected to resume using it. However, detailed research on this user's return to the service has not yet been conducted. In this study, we focused on "returnee users" and analyzed the buyers’ behavioral data at minne, a handmade C2C e-commerce. This study clarifies the significance and the way of promoting their return. First, it is shown that returnee users tend to purchase more times than new users, and acquiring them has a high effect on service revitalization. In addition, it is shown that the psychological hurdle for returning increases as the withdrawal period becomes longer, and goods for hobbies, or those which can be easily purchased from the other competitors become insufficient to motivate users to return. On the other hand, products related to life events or anniversaries, and products that the service is considered special about, may motivate them to return. Furthermore, it became clear that recommending new products from producers whose products they purchased in the past is effective to encourage the users to return to the service.

In the theory of consumer behavior, it is known that purchasing behavior types can be classified according to the level of involvement in a product category and the perceived differences between brands of consumers respectively. In this paper, we investigate the applicability of the purchasing behavior categories to Internet-based services and the differences among service categories. Specifically, we analyzed four types of Internet services, namely "paid video services," "subscription music services," "e-commerce sites," and "consumer-generated media," by questionnaire survey about their involvement levels and perceived differences among brands. As a result, it was found that involvement levels and perceived differences differed among clusters of consumers based on clustering according to the service used across categories, rather than within each service category.

In order to improve household finances, it is important to understand one's own lifestyle. For example, by understanding that one uses convenience stores frequently, one could partially switch to using supermarkets in order to reduce one's household expenditures. In this study, we attempt to predict the user's lifestyle from household account book data using a topic model. In the future, we expect to apply the result of this study to provide appropriate information and suggest ways to improve user's household finances depending on each user's lifestyle.

In recent years, acquisition of strategies has been successfully achieved in video games as well as board games by using self-play. In this research, we report on a study of strategy learning in single player and competitive falling-puzzle game Puyo-Puyo using self-play and deep reinforcement learning. Self-Play is a method in which agents play against each other. In this experiment, we created a puzzle game environment using Unity and ML-Agents and trained using the deep reinforcement learning algorithm SAC. The single player Puyo-Puyo was evaluated on cumulative rewards and maximum number of chains. Although there was a temporary improvement in performance, the result was a little worse. In the competitive Puyo-Puyo was evaluated on Elo-Rating and maximum number of chains. Elo-Rating increased from 1200 to 3100 and it was on an upward trend. It is possible that future studies will make it stronger.

Recently, the application of deep reinforcement learning to game environments has attracted attention. In particular, the game with imperfect information has been paid attention to in this field. In this research, we focus on the trading card game (TCG). TCG is more difficult to be played with artificial intelligence than other games because the performance and types of available cards can be changed. This nature also makes it difficult to adjust the game balance, and it is common for the game to be modified after its release, and terms such as "buff" to change the card performance upward and "nerf" to change it downward are used. Based on the above background, we propose game balance optimization methods for TCG environments using deep reinforcement learning and evolutionary computation and demonstrate the effectiveness of the proposed methods through numerical experiments using our own TCG environment.

The development of 3D contents such as games and movies is becoming larger and larger every year, and the cost of producing motion data for 3D characters accounts for a large percentage of the total cost. In order to streamline the production process, it is necessary to search for similar 3D motions. In this study, we propose a self-supervised Vision Transformer (ViT) for 3D motion retrieval. Using 3D motion data from actual game products, we compare the proposed method with existing CNN-based methods and supervised ViT, and show the improvement in accuracy. We also developed a 3D motion retrieval web system using the proposed method. In this study, we obtained feedback from motion creators who used the Web system in their actual game development flow and discuss it.

The development of AI technology has led to the development of game AI that outperforms professionals. However, because it is too strong, only a few people have been able to take advantage of it. In this study, we propose a rival AI method that dynamically adjusts the strength of the AI so that it becomes as strong as its opponent, and allows a strong AI to be used as a training partner. In this study, we have implemented the proposed method on a strong Shogi AI created by supervised learning using a game record. In the experiment, the Shogi AI was able to make appropriate moves against an amateur 5-kyu level Shogi player and an amateur 1-dan level human player.

Soccer is one of the least constrained and most complex team sports, which makes it extremely hard to capture its behavior. In recent years, attempts have been made, mainly using machine learning techniques, to predict event sequences that indicate which player has taken what action where in a soccer match. However, the prediction of an event seems to depend not only on the previous event sequences, but also to a large extent on which player performs the action. In this study, we propose leveraging a distributed representation, i.e. a vector representation of the players as input for a neural event prediction model. In this way, the model can take into account player characteristics that have not been considered in previous studies. Our results demonstrated that "player vectors" improved the performance of neural event prediction models and that these vectors contain relevant information about player roles.

The focus of this research is on the cooperative connected autonomous vehicles (cooperative CAVs). The aim of this research is to propose a simulator and a environment to enable the learning and implementation of cooperative CAVs. Since there are many challenges to implement cooperative CAVs, large-scale demonstration have not yet been conducted. Therefore, it is valuable to conduct experiments of cooperative CAVs by large-scale simulations. In this research, we propose a simulator and a environment to learn cooperative CAVs driving policy using RL. A communication feature is incorporated to the simulator by sharing the observation information determined based on distance in a chain. Experiments are conducted to determine if cooperative CAVs can be implemented by sharing the observation information. We compared the rewards obtained by learning with and without sharing the observation information. Results show that cooperative CAVs cannot be implemented solely by sharing the observation information, as the reward obtained is higher when the observation information is not shared. Through discussions, we identified issues that need be addressed to implement cooperative CAVs.

Traffic light recognition is critical to the realization and social implementation of self-driving cars. Traffic lights in Japan generally consist of three colors (red, yellow and green) and are often oriented horizontally, but the shape, size, and arrangement of colors vary depending on the country and city. With the advancement of image processing technology, traffic light recognition using deep learning-based object detection models has been researched. However, most large-scale datasets were collected overseas, and only a few were managed and constructed in Japan. Furthermore, it is still challenging to improve the accuracy of the object detection models and understand the context of recognized traffic lights and travel lanes. In this study, we collected more than 900 hours of driving data collected independently on Japanese public roads and constructed a dataset of 15,000 images of multiple scenes annotated with traffic lights. Using this dataset, we trained a deep learning-based 2D object detection model to recognize the traffic lights corresponding to the own driving lane. We verify that the proposed model is specialized to Japanese traffic data and is helpful for traffic light recognition in self-driving cars.

In this work, we propose a Trimodal Navigable Region Segmentation Model (TNRSM), which can handle three modalities: image, navigation instruction, and semantic segmentation masks. We validated our model on the Talk2Car-Regseg dataset. The results demonstrated that our method outperformed the baseline method in all evaluation metrics in the Referring Navigable Regions task.

Due to the expansion of the E-commerce market, the impact of the COVID-19 infection, and the lack of delivery resources caused by the recent labor shortage, the importance of the home delivery business and the need to improve delivery efficiency are increasing in various industries. In particular, the last-one mile, which is the part of delivery from the final location such as a store to the consumers, is facing the need to consider a variety of new requirements and constraints due to the increasing demand for immediate delivery (Q-commerce) in addition to conventional E-commerce delivery. In this study, we develop a delivery optimization algorithms for both "immediate delivery" and "planned delivery" as part of system construction for a delivery platform to address the above-mentioned issues. Specifically, we developed a driver matching optimization algorithm for immediate delivery and a route optimization algorithm for planned delivery. In both cases, we comprehensively took into account the optimization requirements and constraints necessary to solve problems in practical operations, and conducted comparisons and experiments of appropriate algorithms and models. We were able to observe a constant rate of improvement in various evaluation indices by applying the algorithms to actual business operations.

Due to recent outbreaks of infectious diseases and miscellaneous accidents, it has become important to understand the state of crowds. In this study, we used Channel State Information (CSI), which represents the transmission status of Wi-Fi radio waves, to estimate the location of people in a room by considering the amplitude of each subcarrier as a feature as a multi-level classification task in a random forest. In laboratory experiments, the accuracy was high when the training data include the data of the person whose location was to be estimated, but the accuracy deteriorated when the training data did not. In addition, the use of moving average of the series data improved the estimation accuracy for many combinations of data.

Accurate simulation of physical systems is required in various fields of real-world systems. To automatically build a model from the data, recent studies have attempted to use deep learning to build models of the system. Neural ordinary differential equation (Neural ODE), which treats the output of a neural network as the time derivative of the input, has brought development to this research field. However, the training strategy of Neural ODE and related methods still needs to be established. We proposed the error-analytic strategy as a new strategy for training time series datasets to be more accurate in long-term predictions. The proposed strategy is inspired by error analysis techniques in numerical analysis and is derived by replacing numerical errors with modeling errors. Our strategy can capture a long-term error and hence improve the performance of long-term predictions.

This paper discusses the task of anomaly detection and localization from audit data. In auditing, anomaly detection is required in many situations and there exists a strong need for a method that automates the anomaly detection process. However, it is not trivial how to construct audit anomaly detection method due to three difficulties: (1) the method should be unsupervised as it is difficult to manually assign labels to large amounts of data; (2) it is required to conduct anomaly detection and localization simultaneously; (3) an audit data includes both categorical and numerical variables and they correlate strongly. We propose an audit anomaly detection method which solves the above-mentioned difficulties. The key ideas of the method are: (1) we decompose the anomaly detection problem into multiple scenarios, which consist of a few variables, and each scenario corresponds to localization; (2) we unify the localization scenarios by unsupervised ensemble learning which we propose here. We demonstrate effectiveness of the proposed method through the experimental results using anonymized audit data.

This study proposes a new method for constructing statistical arbitrage strategies. Statistical arbitrage is an investment strategy derived from arbitrage. Both strategies exploit distorted relationships between asset prices. However, they differ in that arbitrage assumes a deterministic adjustment of distortions between asset prices, whereas statistical arbitrage assumes a stochastic adjustment of distortions. Therefore, the critical issue for trading strategies that target statistical arbitrage is whether we can stably operate the statistical arbitrage strategy. However, methods based on cointegration tests, which have often been used in previous studies of statistical arbitrage, do not necessarily construct stable portfolios. The main objective of this study is to overcome such problems. Specifically, we attempt to find stable portfolio components with the help of eigenvalue decomposition and improve the profitability and stability of portfolios using standard deep learning architectures such as GRUs and CNNs. Through numerical simulations using a dataset of stocks listed in the S&P 500, we first demonstrate the superiority of our method over conventional methods based on cointegration tests. Then, we verify that the method's performance can be improved with the help of deep learning.

To mitigate the damage caused by malware, network log analysis with machine learning for detecting suspicious logs has been attracting attention. In actual security operation, the true positive rate (TPR) in a low false positive rate (FPR) is important since operators must detect as many suspicious logs as possible while suppressing false positives. This paper focuses on the partial area under the curve (pAUC) maximization method that directly maximize the TPR in an arbitrary FPR interval. However, when using the previous pAUC maximization methods in actual operation, the classifier prone to overfitting and the classification performance tends to be deteriorate if there are mislabelings in the training data. To solve the problems, we propose the method that combines the AUC maximization and pAUC maximization method according to the mathematical characteristics of the features. We also demonstrate the effective of proposed method with proxy logs from a real-world large enterprise network.

In this study, we propose a new approach to intellectual property management to enhance knowledge discovery by combining causal extraction and similarity analysis. Existing research is limited to mining patent text data and predicting technology trends in the same field. Our proposed method can broaden the potential application of technology in multiple areas and enable the secondary development of the intellectual property. Specifically, we implement an approach to extract information from patent text data in various domains. Then, patent text analysis is performed using a pre-trained language model to compare information, and ultimately a causal chain for knowledge discovery is constructed. Data based on expert evaluation demonstrates that our method is more robust than existing deep learning methods.

Hybrid AI of credit models refers to advanced AI credit systems by combining different types of AI techniques and theories and DB techniques, such as machine learning and Bayesian statistics, machine learning of unbalanced data or alternative data. We will discuss the simultaneous tuning of initial and developing credit models, factoring models and credit model trends in the construction industry created by Hybrid AI.

Protein structures constantly fluctuate, and dynamics specific to each time scale are known. In particular, time scales from microseconds to seconds are indispensable for obtaining a complete picture of structural dynamics. Several methods have been proposed to represent slow structural dynamics in low dimensions. However, the existing methods cannot capture such slow dynamics that rarely occur. We propose a method that introduces a prior distribution with high autocorrelation so that even rare slow changes can be emphasized. The proposed method can capture even rare slow dynamics in the representation space by promoting sample-wise autocorrelation. Applying the proposed method to simulated protein trajectories shows that the proposed method can represent slow structural dynamics.

Molecular crystals are composed of organic molecules, and molecular and crystal structures can be used in materials informatics (MI). There is a scarce report on the effectiveness comparison of molecular and crystal representations on the prediction task, and a fundamental question arises in MI of molecular crystals: which descriptor is useful and how effective it is in prediction task. This work aims to answer this question We used bandgap as the target property. Using this dataset, we compared regression results of molecular and crystal graphs through some graph neural network. As the result, molecular graph afforded higher prediction accuracy for bandgap, and we discussed the possible reason.

If the physical property values of materials can be evaluated by AI, it is expected that material search can be greatly accelerated. Aiming to explore new battery materials, we will create appropriate training data from academic papers, and construct an ionic conductivity prediction model. In this study, we constructed a Li-ion conductivity prediction model for perovskite-type solid electrolyte. In creating training data, we found the following three problems: (1) Errors in extracting data from papers, (2) Errors by the authors of the paper (measurement errors, incorrect citation of results in reference papers, etc.), (3) Mixing of data with different experimental conditions. (1) can be solved to some extent by improving our data extraction process. On the other hand, for (2) and (3), technology to detect inappropriate data is required. In this study, even when using the training data that was prepared with sufficient care for (1), the correlation coefficient between the predicted value and teacher data was only about 0.5. Therefore, we developed a technique to exclude inappropriate data using machine learning and succeeded in improving the correlation coefficient to 0.84.

In industries, process monitoring is critical for realizing efficient and safe production. The scale of industrial processes has become larger and larger and more and more complicated, and various process monitoring systems, such as multivariate statistical process control (MSPC), have been widely adopted. MSPC monitors process conditions based on the correlation among the process variables by extracting features from the highly correlated and high-dimensional process data. Principal component analysis (PCA) is a popular linear dimensionality reduction method and utilized in MSPC although it is difficult for PCA to deal with the nonlinearity and the process dynamics. In this study, a new process monitoring method based on the nearest correlation (NC) method and variational graph auto-encoder (VGAE) is proposed to cope with the process nonlinearity and dynamics. This study reported the results of its application to the vinyl acetate monomer production process and compared it with conventional process monitoring methods.

This work aims to construct a classification model based on embeddings of multivariate time series in order to extract features more effectively from a dataset with a mixture of time-series and stationary data. The model is based on the Graph Deviation Network (GDN) and learns an embedding representation of each time-series attribute while considering the interdependence of multiple time-series attributes. The GDN-applied model is constructed as an End-to-End model with transfer learning in the pretraining and fine-tuning framework in order to "learn embedded representations of time-series attributes" and "learn a classifier considering stationary attributes" in a series of learning processes. In comparison with several conventional methods, the proposed GDN-applied model outperforms the conventional method (gradient boosting) in prediction accuracy, demonstrating the usefulness of the GDN-applied model. In an additional experiment with class imbalance, replacing the loss function of the GDN-applied model from Cross Entropy to Dice Loss resulted in an improvement of effectiveness, indicating that the application of Dice Loss to the GDN-applied model is effective in dealing with class imbalance.

Preventing rapidly and largely falling market prices is significantly important to prevent financial crisis, so some stock exchanges implement them. There are many discussions for the question which regulation prevents rapidly and largely varying more effectively. In this study, I investigated which is more effective to prevent falling market prices using the artificial market model which is an agent-based model for a financial market. In the result, the both basically prevent to fall almost same effectively when those of parameters, limit price range and limit time range, are same. However, the price limit is poorer effective than the circuit breaker when limit time range is smaller than cancel time range. In the case with the price limit many sell orders are accumulated around the lower limit price. Here, the lower limit price is changed before the accumulated sell orders are cancelled, and it leads to make more accumulated sell orders on various prices. Such the accumulated sell orders on various prices play a role like as wall against buy orders, and the wall prevent to rise prices by some buy orders. It should be taken very carefully that the result of this study showed in the limited situation. The result, the circuit breaker is better than the price limit, is adapted only in the case that the reason why prices are falling is erroneous orders and that individual stocks are regulated.

Aucnet, Inc. provides a used car appraisal support service for used car dealers in Japan. Prior to 2020, prices were manually determined by experienced staff based on car information such as model, mileage, and body conditions. However, speed and accuracy are desired to improve due to increasing size of inquiry. In this study, we developed a price recommendation system using AutoML and compared its performance with human assessment. We used empirical data from Aucnet used car auctions to build a machine learning model that predicts winning bids or prices from precise car conditions. Among the models generated by AutoML, LightGBM performed best with root mean squared errors. We built a system on databricks and put it into actual operation in 2020, resulting in a 28.3\% reduction in response time on average over 100 thousand inquiries. The system also accurately predicted actual winning bids of subsequent auctions, with 73.1\% of the data falling within plus or minus 10\% of the price range suggested by the system.

In the manufacturing industry, there is a problem that it is difficult to predict with sufficient accuracy for practical use because the amount of training data is small with only the data of the prediction target product. Therefore, in this method, in the extraction of training data from the source domain, the training data of the target domain is extended by performing clustering using time-series features that represent the demand characteristics of electronic devices. We compared our method with conventional methods that do not use domain adaptation, and improved the short-term demand forecast accuracy.

To reduce food waste, supermarkets discount the price of food that is nearing its expiration date. To maximize profits, it is necessary to adjust the discount rate as low as possible while controlling waste due to unsold products. Traditionally, the operational discount rate has been determined based on the experience of each store operator. Inexperienced operators have difficulty in determining the appropriate discount rate for the situation, resulting in excessive discounting and disposal losses. In this study, we developed an algorithm that suggests the optimal discount rate to maximize store profits. Specifically, we developed a demand forecasting model that predicts the expected sales volume per hour for each product based on POS data, etc., and a discount optimization algorithm that uses dynamic programming to calculate the optimal discount rate for each product, time, and number of items in stock based on the forecast results. We conducted a validation experiment of the algorithm in a store, taking into account the constraints of real operations, and observed a constant rate of improvement in discounting and waste loss.

The advertising budget allocation problem is the problem of determining when, to which advertising media, and how much to allocate the budget to maximize advertising effectiveness. This problem changes its characteristics depending on the advertising media assumed. In this paper, we consider a setting that includes both direct buying advertising, such as TV, and programmatic advertising, such as the Internet. We formulate the problem as a Markov decision process and apply model-free reinforcement learning. We demonstrate the effectiveness of our approach by experimenting with real-world data.

The performance of a ship is greatly affected by external forces such as wind, waves and conditions of the ship such as hull fouling. Since the effects of hull fouling can be reduced by ship maintenance, conversely quantitative analysis of impact of hull fouling can be a good approach to determine the maintenance at the optimum timing and specifications. However, it is difficult to estimate the impact of hull fouling directly because multiple factors intricately affect external forces and the condition of the ship. This paper proposes a model that estimates the impact of hull fouling by using a neural network that predicts the speed of the ship from external forces such as wind and waves. The experimental results of three ships demonstrated the strong relationship between the impact of hull fouling and the effects of maintenance in a time series.

Various studies have been conducted on automatic ship maneuvering in bays, but few studies have specifically addressed obstacle avoidance. This study aims to develop a control law that combines obstacle avoidance and docking control using reinforcement learning. Since the location of obstacles is uncertain during policy training, it is necessary to generalize the policies for different obstacle configurations. This work proposes a method that generates the distribution of the initial states of the obstacles, which is used in domain randomization to obtain a berthing control law that can handle obstacles in unspecified locations and achieve berthing control at the target state. Simulation results show that the proposed approach has a high success rate in avoiding obstacles and achieving berthing control for the ship in the majority of trials.