CS294-112 Final Project

Bridging Distribution Mismatch: Better Bound on Composable Deep Reinforcement Learning

Dependencies:

• Python 3.5
• Numpy version 1.14.5
• TensorFlow version 1.10.5
• OpenAI Gym version 0.10.5
• seaborn
• Box2D==2.3.2
• OpenCV
• ffmpeg
• Vizdoom

Before doing anything, first replace gym/envs/box2d/lunar_lander.py with the provided lunar_lander.py file.

The code was based on an implementation of Q-learning in HW3.

This code can support Atari games and Vizdoom environments.

Instructions for Composable Soft Q Learning

python run_dqn_vizdoom.py --vizdoom --explore e-greedy --subgame shoot_monster
python run_dqn_vizdoom.py --vizdoom --explore soft_q --subgame avoid_shooters
python run_dqn_vizdoom.py --vizdoom --explore soft_q --ex2 --coef 1e-4 --subgame avoid_shooters

Cheers!