DeepMine - Survey of different flavours of Deep Reinforcement Learning

In this project we chose 6 most recent deep reinforcement learning algorithms, made an experience using Open AI's Lunar Lander environment and analyzed the result.

How to run?

1. DQN

cd Open-AI/DQN2

To train the model (takes too long: ~1,5 days)

python3 dqn.py 0

To load the model add model number as an extra parameter. Model is already trained and output files are commited. So if you want to load final model with parameter 0: python3 dqn.py 0 0

In each 100 episode model is saved with '.h5' extension. So in order to load model after 30000 episodes: python3 dqn.py 0 300

2. Prioritized Experience Replay

cd Open-AI/DQN-PER

All the other part is same with DQN's running.

3. Policy Gradient

To train the model

cd Open-AI/PG

python3 0.001_loss_steps.py.py

4. A3C

To train the model

cd Open-AI/A3C

python3 raw_A3C.py