fangzai / atari-deep-reinforcement-learning Goto Github PK

##Deep Reinforcement learning to play Atari games

This project contains the source code of DeepMind's deep reinforcement learning architecture described in the paper "Human-level control through deep reinforcement learning", Nature 518, 529–533 (26 February 2015).

The following changes to DeepMind code were made:

• [cuDNN support] (https://developer.nvidia.com/cudnn)

• Different weight initialization methods:

  • Xavier Glorot, Yoshua Bengio "Understanding the difficulty of training deep feedforward neural networks", 2010.
  • Kaiming He, et al. "Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification", 2015

• Reward normalization within minibatch

• Independent subspace analysis (ISA) can be used as preprocessing step. In order to apply it, change agent name in run_gpu to NeuralQLearnerISA (agent="NeuralQLearnerISA").

####Installation

To install all dependencies, it should be enough to run

./install_dependencies.sh

Alternatively, you can use following AWS ami with preinstalled dependencies: ami-b36981d8

Prior to running DQN on a game, you should copy its ROM in the 'roms' subdirectory. It should then be sufficient to run the script

./run_gpu <game name>