Using Deep Neural Network Based on Reinforcement Learning for Video Game Control Strategy

• Main Authors: Lu-Xun Huang, 黃祿勳
• Other Authors: Cheng-Yi Yu
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/kn92fw

碩士 === 國立勤益科技大學 === 資訊工程系 === 106 === This paper proposes a game control method that uses game screens as input, improves the architecture of the roll machine neural network, initializes the parameters of the convolution kernel using principal component analysis, and stack these convolution kernels through multiple layers. Learning how to extract feature from input image and base on these feature to make control strategy by reinforcement learning method, effectively realize the game control method that imitates humans using vision as input. This architecture is named Reinforcement Q-Learning on Deep Neural Network (RQDNN). In the experiments of this paper, RQDNN is better than human players and other deep reinforcement learning algorithms. In addition, the RQDNN is much less than the other deep reinforcement learning algorithms using convolutional neural networks for the computational resources required for training.