rvdweerd / rl-experiments

Reinforcement Learning experiments, comparing performance of Q-learning and Double Q-learning algorithms.

This project compares the Vanilla Q-learning and Double Q-learning algorithms. Grid worlds (inherited classes from gym.Env with 4-dimensional action space NWSE) can be defined with various cell types:

• Lava: reward -100 and episode terminates
• Cliff: reward -100, episode continues from start position
• Wall: reward -1, no position change
• Regular: reward -1, regular position change

And various types of stochasticity:

• State transition stochasticity: defines some probability of veering off to left/right when moving forward
• Reward stochasticity: reward per step is [-12,+10] with equal probability (instead of deterministic -1)

Example experimental setups are in ./configs. Required Python packages are in environment.yaml. To reproduce results, run

python main.py [CONFIG_NAME.yaml]

Example experiment

Here are some results (200 seeds, 300 episodes per seed) for a simple maze with walls (black), lava (orange) and cliff (grey) cells. Start position in blue, goal position in green.