Atari RL

This project implements reinforcement learning agents based on Deep Q-Networks (DQN) and Double DQN for playing Atari 2600 games using the OpenAI Gym environment. It includes key features such as experience replay, target networks, ε-greedy exploration, and support for multiple Atari titles.

Create conda environment

mamba env create -f environment.yaml

Run local mlflow

mlflow ui

Train using params profile

# for pytorch deterministic algorithms
export CUBLAS_WORKSPACE_CONFIG=:4096:8
python dqn_train.py --profile double-dqn-tuned-adam

Evaluate model stored in mlflow

python dqn_eval.py --profile double-dqn-tuned-adam --model_uri 'runs:/8d3c31e65a3240eda1e0e92890057f46/q0_episode_22000'

Alternatives that train faster

• Rainbow DQN (2017): Uses prioritized replay, multi-step returns, and distributional RL to speed up learning.
• Ape-X DQN: Uses multiple learners for faster wall-clock time.