README.md

RL289A-WQ2020

Attempting to Solve Sokoban using DQN

Final project for EEC 289A Reinforcement Learning Course

Group Members

• Kolin Guo
• Daniel Vallejo
• Fengqiao Yang

Prerequisites

• Ubuntu 18.04
• NVIDIA GPU with CUDA version >= 10.1
• Docker version >= 19.03, API >= 1.40
• nvidia-container-toolkit (previously known as nvidia-docker)

Command to test if all prerequisites are met:
sudo docker run -it --rm --gpus all ubuntu nvidia-smi

Setup Instructions

bash ./setup.sh
You should be greeted by the Docker container openaigym when this script finishes. The working directory is / and the repo is mounted at /RL289A-WQ2020.

Running Instructions

• Training from scratch
  python3 src/train.py
  Resume training from a checkpoint file
  python3 src/train.py --checkpoint_dir checkpoints/DQN_Train --checkpoint_file ckpt-100000
• Testing
  python3 src/test.py --checkpoint_dir checkpoints/DQN_Train
• Playing (generating game-play examples using training checkpoints)
  python3 src/play.py --checkpoint_dir checkpoints/DQN_Train --checkpoint_file ckpt-100000

Some other available arguments can be viewed with --help option.

Presentation and Report

Our final presentation (with embedded audio) and report can be found in docs/ folder.
Some additional improvements (CNN+LSTM model, deadlock detection algorithm, A3C algorithm) are discussed at the end of our presentation.