abhishm / dqn

A short tutorial about tips and tricks used in implementation of Deep Q-networks that helped in creating a first AI that can successfully play Atari.

DQN

Modular implementation of DQN algorithm.

Dependencies

• Python 2.7 or 3.5
• TensorFlow 1.10
• gym
• numpy
• tqdm progress-bar

Features

• Using a neural network based as the function approximator for Q-learning
• Using a target network and soft-update to synchronoze target network with Q-network
• Using gradient clipping to make small but consistent updates towards optimal Q-network

Bonus

• Implementation of Tabular Q-learning

Usage

To train a model for Cartpole-v0:

$ python test_graph_dqn.py 

To view the tensorboard

$tensorboard --logdir .

Results

• Tensorboard Progress Bar