Self-learning game bot using deep reinforcement learning

Abstract

We present a deep learning model for playing games with high level input (image/raw pixel) using reinforcement learning. The games are action limited (like snakes, catcher, air-raider etc.). The model consists of convolution neural network for processing image inputs and fully connected layers for estimating actions according to the inputs where the idea of taking action is based on Q-learning (model-free reinforcement learning), yet modified it for our policy and usage. We applied our method on the python’s ‘PyGame Learning Environment’ and some other classic control games. We found our method learns fast enough but not with best accuracy. Then we tried the batch of input method which results a high score for the Catcher environment. It produced better performance in terms of the learning speed and accuracy.