Search

Mahjong, a complex game with hidden information and sparse rewards, poses significant challenges. Existing Mahjong AIs require substantial hardware resources and extensive datasets to enhance AI capabilities. The authors propose a transformer‐based Mahjong AI (Tjong) via hierarchical decision‐making. By utilising self‐attention mechanisms, Tjong effectively captures tile patterns and game dynamics, and it decouples the decision process into two distinct stages: action decision and tile decision. This design reduces decision complexity considerably. Additionally, a fan backward technique is proposed to address the sparse rewards by allocating reversed rewards for actions based on winning hands. Tjong consists of 15M parameters and is trained using approximately 0.5 M data over 7 days of supervised learning on a single server with 2 GPUs. The action decision achieved an accuracy of 94.63%, while the claim decision attained 98.55% and the discard decision reached 81.51%. In a tournament format, Tjong outperformed AIs (CNN, MLP, RNN, ResNet, VIT), achieving scores up to 230% higher than its opponents. Furthermore, after 3 days of reinforcement learning training, it ranked within the top 1% on the leaderboard on the Botzone platform. [ABSTRACT FROM AUTHOR]

Automated decision-making systems have been deployed in many industrial, commercial, and financial applications. The needs for such systems are usually motivated by requirements for variation reduction, capacity increase, cost and cycle time reduction, and end-to-end traceability of the transaction or product. Before we can use any automated decision-making system in a production environment we must develop a strategy to insure high quality throughout its entire lifecycle. We need to guarantee its performance through a rigorous Design for Six Sigma process (DFSS). This process includes validation, tuning, and production testing of the system. Once the system is in production we must monitor and maintain its performance over its lifecycle. In this paper we will outline the Six Sigma process that led to the deployment of an automated decision-making system in one of the General Electric Financial Assurance businesses. Copyright © 2005 John Wiley & Sons, Ltd. [ABSTRACT FROM AUTHOR]

Despite much progress in brain and cognitive sciences, attempts to connect brain function to cognition are hampered by the large explanatory gap between psychology and neurobiology. In recent years, a neurocomputational perspective has emerged as the most promising approach to integrating brain and mind. According to this perspective, the brain is a special sort of computer, a system of many parallel neural networks whose operation underlies cognition. In this paper, we present this neurocomputational perspective and examine the ways in which this new approach to explaining our mental skills differs from earlier ones. In particular, we examine its emerging insights into two domains. First, we explore the neurocomputational approach to decision-making, the adaptive guidance of behavior in the satisfaction of life maintenance goals. Decision-making is central to all mobile creatures in an uncertain environment, and this approach reveals a surprising conservation of decision-making strategies across many species. We then examine the neurocomputational approach's new insights into characterizing cognitive development. In particular, this approach offers the new framework of self-organization to characterize the complex interaction between neural developmental programs and the environment, a framework that has important implications for understanding early intervention. MRDD Research Reviews 1999;5:86–99. © 1999 Wiley-Liss, Inc. [ABSTRACT FROM AUTHOR]

Objectives: This study aimed to develop a diagnostic support tool using pretrained models for classifying panoramic images of the temporomandibular joint (TMJ) into normal and osteoarthritis (OA) cases. Subjects and Methods: A total of 858 panoramic images of the TMJ (395 normal and 463 TMJ‐OA) were obtained from 518 individuals from January 2015 to December 2018. The data were randomly divided into training, validation, and testing sets (6:2:2). We used pretrained Resnet152 and EfficientNet‐B7 as transfer learning models. The accuracy, specificity, sensitivity, area under the curve, and gradient‐weighted class activation mapping (grad‐CAM) of both trained models were evaluated. The performances of the trained models were compared to that of dentists (both TMD specialists and general dentists). Results: The classification accuracies of ResNet‐152 and EfficientNet‐B7 were 0.87 and 0.88, respectively. The trained models exhibited the highest accuracy in OA classification. In the grad‐CAM analysis, the trained models focused on specific areas in osteoarthritis images where erosion or osteophyte were observed. Conclusions: The artificial intelligence model improved the diagnostic power of TMJ‐OA when trained with two‐dimensional panoramic condyle images and can be effectively applied by dentists as a screening diagnostic tool for TMJ‐OA. [ABSTRACT FROM AUTHOR]

Context Clinical supervisors make judgements about how much to trust learners with critical activities in patient care. Such decisions mediate trainees' opportunities for learning and competency development and thus are a critical component of education. As educators apply entrustment frameworks to assessment, it is important to determine how narrative feedback reflecting entrustment may also address learners' educational needs. Methods: In this study, we used artificial intelligence (AI) and natural language processing (NLP) to identify characteristics of feedback tied to supervisors' entrustment decisions during direct observation encounters of clerkship medical students (3328 unique observations). Supervisors conducted observations of students and collaborated with them to complete an entrustment‐based assessment in which they documented narrative feedback and assigned an entrustment rating. We trained a deep neural network (DNN) to predict entrustment levels from the narrative data and developed an explainable AI protocol to uncover the latent thematic features the DNN used to make its prediction. Results: We found that entrustment levels were associated with level of detail (specific steps for performing clinical tasks), feedback type (constructive versus reinforcing) and task type (procedural versus cognitive). In justifying both high and low levels of entrustment, supervisors detailed concrete steps that trainees performed (or did not yet perform) competently. Conclusions: Framing our results in the factors previously identified as influencing entrustment, we find a focus on performance details related to trainees' clinical competency as opposed to nonspecific feedback on trainee qualities. The entrustment framework reflected in feedback appeared to guide specific goal‐setting, combined with details necessary to reach those goals. Our NLP methodology can also serve as a starting point for future work on entrustment and feedback as similar assessment datasets accumulate. Using natural language processing, Gin et al demonstrate how narrative feedback and entrustment decisions are intertwined, providing empirical evidence regarding how formative processes can help trainees achieve competencies. [ABSTRACT FROM AUTHOR]

This paper proposes the Actor-Q architecture, which is a combination of Q-Learning and Actor-Critic architecture, as well as the active perception and recognition learning system based on that architecture. In Actor-Q architecture, the system output is divided into the “action,” which is a discrete value as intention, and the “motion,” which is a continuous-valued vector. As the first step, the “action” is determined from Q-values. If the “action” is accompanied with a “motion,” the “motion” is executed according to the corresponding Actor output. Q-value is learned by Q-learning, and Actor is trained with the Q-value corresponding to that “action” on behalf of the Critic output. In this study, the action is defined as the decision of the sensor motion or the recognition of the respective pattern. Q-value is assigned to each of those. When the sensor motion is selected, the sensor is moved according to the Actor output. When recognition is selected, the recognition result that the presented pattern is the one corresponding to the selected Q-value is output. The Q-value is learned, using the reinforcement signal representing the true/false of the result. Both Q-value computing module and Actor are composed of neural networks, with the visual sensor signals as input. By this architecture, the following three problems of the conventional active perception and recognition learning system are dissolved. (1) The sensor can be trapped in a local maximum of the recognition evaluation. (2) It is necessary that the recognition output should be evaluated for each time-step, and the reinforcement signal with a continuous value should be provided. (3) The system cannot decide by itself the timing to output the recognition result. The above effect was verified by some simulations, using the visual sensor with nonuniform sensor cells. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(14): 12–22, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10207 [ABSTRACT FROM AUTHOR]

This study proposes the use of a modified genetic algorithm (MGA), a global search technique, as a training method to improve generalizability and to identify relevant inputs in a neural network (NN) model. Generalizability refers to the NN model's ability to perform well on exemplars (observations) that were not used during training (out-of-sample); improved generalizability enhances NN's acceptability as a valid decision-support tool. The MGA improves generalizability by setting unnecessary weights (or connections) to zero and by eliminating these weights. Because the eliminated weights have no further impact on the training (in-sample or out-of-sample data), the relevant variables can be identified from the model. By eliminating unnecessary weights, the MGA is able to search and find a parsimonious model that generalizes well. Unlike the traditional NN, the MGA identifies the model variables that contribute to an outcome, helping decision makers to rationalize output and accept results with greater confidence. The study uses real-life data to demonstrate the use of MGA. [ABSTRACT FROM AUTHOR]

A number of recent studies have compared the performance of neural networks (NNs) to a variety of statistical techniques for the classification problem in discriminant analysis. The empirical results of these comparative studies indicate that while NNs often outperform the more traditional statistical approaches to classification, this is not always the case. Thus, decision makers interested in solving classification problems are left in a quandary as to what tool to use on a particular data set. We present a new approach to solving classification problems by combining the predictions of a well-known statistical tool with those of an NN to create composite predictions that are more accurate than either of the individual techniques used in isolation. [ABSTRACT FROM AUTHOR]