Toly Chen’s research while affiliated with National Yang Ming Chiao Tung University and other places

Several artificial intelligence (AI) technologies have been applied to assist in the selection of suitable three-dimensional (3D) printing facilities in ubiquitous manufacturing (UM). However, AI applications in this field may not be easily understood or communicated with, especially for decision-makers without relevant background knowledge, hindering the widespread acceptance of such applications. Explainable AI (XAI) has been proposed to address this problem. This study first reviews existing XAI techniques to explain AI applications in selecting suitable 3D printing facilities in UM. This study addresses the deficiencies of existing XAI applications by proposing four new XAI techniques: (1) a gradient bar chart with baseline, (2) a group gradient bar chart, (3) a manually adjustable gradient bar chart, and (4) a bidirectional scatterplot. The proposed methodology was applied to a case in the literature to demonstrate its effectiveness. The bidirectional scatterplot results from the experiment demonstrated the suitability of the 3D printing facilities in terms of their proximity. Furthermore, manually adjustable gradient bars increased the effectiveness of the AI application by decision-makers subjectively adjusting the derived weights. Furthermore, only the proposed methodology fulfilled most requirements for an effective XAI tool in this AI application.

Deep neural networks (DNNs) have been applied to predict the cycle times of jobs in manufacturing accurately. However, the prediction mechanism of a DNN is complex and difficult to communicate. This limits its acceptability (or practicability) in real-world applications. An explainable deep-learning approach is proposed to solve this problem in this study. This study proposes a classification and regression tree (CART) to explain the prediction mechanism of a DNN for job cycle time prediction. The predicted value of each branch in the CART is replaced by a fuzzy linear regression (FLR) equation that estimates the cycle time range to compensate for the insufficient explainability. The explainable deep-learning approach has been applied to a real-world study from the literature to evaluate its effectiveness. According to the experimental results, the explainability of the prediction mechanism of the DNN, measured in terms of root mean squared error (RMSE), using the CART was high. In addition, the proposed methodology was able to make local explanations.

Many evolutionary artificial intelligence (AI) technologies have been applied to assist job scheduling in manufacturing. Among them, genetic algorithms (GAs) are one of mainstream methods. However, GA applications in this field may not be easy to understand or communicate, especially to factory workers without relevant background knowledge, preventing widespread acceptance of such applications. To address this problem, the concept of explanatory AI (XAI) has been proposed. This study first reviews existing XAI techniques for explaining GA applications in job scheduling. Based on the review results, the problems faced by existing XAI techniques are summarized. To solve these problems, this study proposes several novel XAI techniques, including decision tree-based interpretation, dynamic transformation and contribution diagrams, and improved bar charts. To illustrate the effectiveness of the proposed methodology, it has been applied to a case in the literature. According to the experimental results, the proposed methodology can make up for the deficiencies of existing XAI methods in processing high-dimensional data and visualizing the contribution of feasible solutions, thereby satisfying all the requirements for an effective XAI technique for explaining GA applications in job scheduling. Furthermore, the proposed methodology can be easily extended to explain other evolutionary AI applications such as ant colony optimization (ACO), particle swarm optimization (PSO), artificial bee colony (ABC) in job scheduling.

Recently, many methods based on artificial neural networks (ANNs) or deep neural networks (DNNs) have been proposed to accurately predict the cycle time of a job. However, the prediction mechanism of an ANN is difficult to understand and communicate for users, which limits its acceptability (or usefulness). To solve this problem, a two-stage explainable artificial intelligence (XAI) approach is proposed in this study to better explain a classification-based cycle time prediction method. In the proposed methodology, first, jobs are divided into several clusters. A scatter radar diagram is then designed to illustrate the classification result. Compared with existing XAI techniques, the scatter radar diagram meets more requirements for better interpretation. Subsequently, an ANN is constructed for each cluster to predict the cycle times of jobs in the cluster. A random forest is then constructed to approximate the prediction mechanism of the ANN. In existing practice, the random forest generates many decision rules to predict the cycle time of a job, which may cause confusion for the user. To solve this problem, a systematic procedure is established to re-organize these decision rules. In this way, the first few decision rules can provide most of the information, and the user does not have to read all the rules. The two-stage XAI approach has been applied to a real case from the literature to evaluate its effectiveness.

In a fuzzy group decision-making task, when decision makers lack consensus, existing methods either ignore this fact or force a decision maker to modify his/her judgment. However, these actions may be unreasonable. In this study, a fuzzy collaborative intelligence approach that seeks the consensus among experts in a novel way is proposed. Fuzzy collaborative intelligence is the application of biologically inspired fuzzy logic to a group task. The proposed methodology is based on the fact that a decision maker must make a choice even if he/she is uncertain. As a result, the decision maker’s fuzzy judgment matrix may not be able to represent his/her judgment. To solve such a problem, the fuzzy judgment matrix of each decision maker is decomposed into several fuzzy judgment submatrices. From the fuzzy judgment submatrices of all decision makers, a consensus can be easily identified. The proposed fuzzy collaborative intelligence approach and several existing methods have been applied to the case of the post-COVID-19 transformation of a Japanese restaurant in Taiwan. Because such transformation was beyond the expectation of the Japanese restaurant, the employees lacked consensus if existing methods were applied to identify their consensus. The proposed methodology solved this problem. The optimal transformation plan involved increasing the distance between tables, erecting screens between tables, and improving air circulation. In a fuzzy group decision-making task, an acceptable decision cannot be made without the consensus among decision makers. Ignoring this or forcing decision makers to modify their preferences is unreasonable. Identifying the consensus among experts from another point of view is a viable treatment.

This study proposes a hybrid big data analytics and Industry 4.0 (BD-I4) approach to enhancing the effectiveness of cycle time range projections for factory jobs. As a joint application of big data analytics and Industry 4.0, the BD-I4 approach is distinct from existing methods in this field. In this approach, each expert first constructs a fuzzy deep neural network to project the cycle time range of a job, an application of big data analytics (i.e., deep learning). Subsequently, the fuzzy weighted intersection operator is applied to aggregate the projected cycle times such that unequal authority levels can be considered, an application of Industry 4.0 (i.e., artificial intelligence). Applying the BD-I4 approach to a real case that the proposed methodology improved the projection precision by up to 72%, suggesting that instead of relying on a single expert, collaboration among multiple experts may be more effective and efficient.

Most of the existing ubiquitous clinic recommendation (UCR) systems adopt linear mechanisms to aggregate the attribute-level performances of a clinic to evaluate the overall performance. However, such linear mechanisms may not be able to explain the choices of all patients. To solve this problem, the modified mixed binary nonlinear programming (MMBNLP)–feedforward neural network (FNN) approach is proposed in this study. In the proposed methodology, first, the existing MBNLP model is modified to improve the successful recommendation rate using a linear recommendation mechanism. Subsequently, an FNN is constructed to fit the relationship between the attribute-level performances of a clinic and its overall performance, thereby providing possible ways to further enhance the recommendation performance. The results of a regional experiment showed that the MMBNLP–FNN approach improved the successful recommendation rate by 30%.