Study of Learning Ability in Profit Sharing Using Convolutional Neural Network

摘要

Profit Sharing using Convolutional Neural Network (PS-CNN) has been proposed as a method of deep reinforcement learning. In the previous work, experiments have been conducted using Atari 2600's Asterix in the Profit Sharing using Convolutional Neural Networks, and it is known that a better score can be obtained than Deep Q-Network. However, experiments have not been conducted on games other than Asterix, and sufficient consideration has not been made. In this paper, we report on the results of studying learning ability for some Atari 2600' games in Profit Sharing using Convolution Neural Network. By comparing the results with the results in Deep Q-Network, we confirmed that this method can acquire higher score than the Deep Q-Network in some games. The common feature of these games is that the number of actions and the number of states are relatively large.