Reinforcement Learning Using Knowledge Controller

Implementation for "Accelerating Deep Reinforcement Learning via Knowledge-Guided Policy Network".

Setup Environment

Environement Requirements

• python 3.7
• pytorch 1.7
• gym 0.16.0
• wandb
• tensorboardX

Our Network Structure and Rules

See Controller in controller.py and XXXRule in ./rule/*.py for details.

Training

All log and snapshot would be stored logging directory. Logging directory is default to be "./output/ENV_NAME". Different environments can be used by providing env values in the args.

# For CartPole without Knowledge
python main.py --env CartPole-v1 --no_controller --max_update 5000 --seed 0

# For CartPole with Knowledge
python main.py --env CartPole-v1 --max_update 5000 --seed 0

# For FlappyBird without Knowledge
python main.py --env FlappyBird --no_controller --max_update 1200 --seed 0

# For FlappyBird with Knowledge
python main.py --env FlappyBird --no_controller --max_update 1200 --seed 0