METHOD FOR SEMI SUPERVISED REINFORCEMENT LEARNING USING DATA WITH LABEL AND DATA WITHOUT LABEL TOGETHER AND APPARATUS USING THE SAME

摘要

The present invention relates to a method for conducting semisupervised reinforcement learning jointly using labeled data and unlabeled data, and a device using the same. To be more specifically, the method allows a computing device to train a baseline neural network as an immediate compensation indicator by using labeled data when acquiring the labeled data and unlabeled data, and to train a policy neural network for searching a subset of the unlabeled data, wherein the subset is searched so that the validation accuracy of the immediate compensation indicator can increase during additional training of the immediate compensation indicator. Then, the computing device additionally trains the immediate compensation indicator by using the label given to the subset through the subset and the policy neural network. Therefore, the semisupervised reinforcement learning method can improve accuracy and efficiency of diagnosis assistance by machine learning.