Pattern recognition techniques are a class of method used to categorize observations across multiple classes. The structure and learning method of a model are selected so as to maximize the classification accuracy obtained over a certain data set. Methods also exist to improve accuracy by suitably structuring the training data [1]. Ensemble learning [2, 3, 4, 5] can be seen as one such method. In this talk, we discuss a new method that improves the structure of data in order to improve the accuracy of a classifier. In contrast to AdaBoost, a conventional boosting method that increases the weight for misclassified data and adds data over learning iterations, we propose Bagging/Boosting methods that generate virtual data using fuzzy membership functions [6] centered on misclassified data. We call this the pdi-Bagging/Boosting method (Possibilistic Data Interpolation Bagging/Boosting method) [1, 7, 8, 9, 10]. The newly generated virtual data around the misclassified data is used exclusively for training - not for validation or testing. Model accuracy is improved by adding virtual data to the original training data because the pattern recognition model is identified by the training data. On the other hand, the accuracy of the identified model is evaluated on the original test data which is not augmented with virtual data. In other words, the virtual data is not involved in the evaluation of the model at all. We apply the pdi-Bagging/Boosting method to Trapezoidal fuzzy inference. Trapezoidal fuzzy inference is a general form of triangular fuzzy reasoning, that has proven to be effective at solving various types of inference problems [11, 12]. When used as a clustering method of bagging/boosting algorithm, fuzzy inference allows efficiently for adjusting cluster boundaries with each new data [13, 14]. In trapezoidal fuzzy inference, both the membership function for the antecedent part and the singleton real value of the consequent part need to be learned [15, 16, 17, 18, 19, 20]. In this talk, we frame the problem of parameter adjustment for fuzzy rules not as a tuning problem but rather, as a design problem [21]. We focus in particular on how to learn coefficients both for the membership functions of the antecedent part and the singletons of the consequent part, how to set initial values, and how to determine the learning sequence of consequent and antecedent parts. We are developing a system for automatically extracting rules of table tennis strategy from video using a clustering method [22, 23]. We analyzed videos from women`s table tennis singles tournament at the 2016 Rio de Janeiro Olympics, including 16 matches from the third round to the final. This video corpus contains a total of 407 plays, and 7,434 separate ball trajectories. In addition, we are collaborating with OMRON Corporation which has been developing a table tennis robot, "FORPHEUS" [24]. In this talk, the possibility of strategy extraction from video or images using ensemble learning algorithm in the near future is discussed, as well as the possibility to characterize players` strategies based on the learned relationship between player characteristics and ball position in image-based coordinates [25]. Through future work, we hope to integrate this new capability into an "AI strategy coach" that can help improve player strategies [26, 27, 28].
© Copyright 2019 IES 2019: Proceedings of the 23rd Asia Pacific Symposium on Intelligent and Evolutionary Systems. Alle Rechte vorbehalten.