Influence on Learning of Various Conditions in Deep Q-Network

As a learning method to acquire an appropriate action sequence by interaction with the environment without using a teacher signal, various researches on reinforcement learning have been carried out. On the other hand, recently, deep learning attracts attention as a method which has performance superior to conventional methods in the field of image recognition and speech recognition. Furthermore, the Deep Q-Network, which is a method can learn the action value in Q-Learning using the convolutional neural network, has been proposed. The Deep Q-Network is applied for many games without adjusting for each game, and it gains higher scores than humans in some games. In this paper, experiments in Deep Q-Network are carried out in the case when the time to be considered as an input is different, and the case when another action selection method is used and so on. As a result, we confirmed as follows: (1) There is a possibility that the performance may be improved by increasing the time to consider as an input to the Deep Q-Network, and (2) There is a high possibility that the probability that an action whose value is maximum in the action selection is chosen influences on learning.