Deep Reinforcement Learning algorithms

This repo contains my implementations of Deep Reinforcement Learning algorithms I used in research:

• DQN (by Deepmind)
• QR-DQN (by Deepmind)
• DLTV (mine)
• DQN-decor (mine)

I tried to make the code concise.

Other features:

1. Experience is collected in a separate thread which gives a slight improvement in runtime.
2. Training reward is logged using tensorboard
3. Hyperparameters (with some exceptions - see the Configs) are taken from Rainbow

How to run

1. setup virtualenv (or conda):

• virtualenv -p python3 p3
• source p3/bin/activate
• pip install -r requirements.txt

2. train agent (see the empirical results in the papers):

• Human-level control through deep reinforcement learning (original DQN)
  • python dqn/agent.py
• Distributional Reinforcement Learning With Quantile Regression
  • python qr-dqn/agent.py
• Distributional Reinforcement Learning for Efficient Exploration
  • python dltv/agent.py
• Deep reinforcement learning with decorrelation
  • python dqn-decor/agent.py

3. check training reward in tensorboard:

• tensorboard --logdir=/tmp/tf_logs