lamikins / dqn

Deep Q-Network and Double Deep Q-Network for Game Playing in Tensorflow

This repository is an implementation of the DQN and Double-DQN from DeepMind

The Deep Q-Network was a seminal deep reinforcement learning paper by Mnih et al., in which the authors demonstrate a system for learning control policies for Atari games directly from visual input.

Usage

Clone the repository, then install all necessary required packages. Training can be started on the OpenAI Gym Cart Pole environment simply by running:

python learn.py

This instantiates a Double Deep Q-Net and trains for 1000 episodes with parameters {image_size: 28x28, learning_rate: 1e-3, sample_time: 1000}

You can modify the hyperparameters in learn.py and change the agent to be either a DoubleDQNLearner or DQNLearner.