Atari Deep Q-Network ProjectAtari Deep Q-Network Project

Overview

This project implements 3 reinforcement learning agents using Deep Q-Networks (DQN) to play:

• CartPole-v1 (classic control)
• SpaceInvadersNoFrameskip-v4
• MsPacmanNoFrameskip-v4

The solution is intended for educational Bootcamp-style demonstration and can be extended for full Atari training on GPU.

Approach

1. Environment setup with OpenAI Gym / Gymnasium.
2. DQN agent training with Stable Baselines3.
3. Atari image preprocessing via AtariWrapper.
4. Periodic evaluation with evaluate_policy.

Running the models

pip install -U gymnasium[atari] stable-baselines3[extra] torch
python atari_cartpole.py
python atari_spaceinvaders.py
python atari_mspacman.py

Results

Expected: CartPole learns to solve within ~100k steps, Atari agents show incremental reward improvements with additional timesteps.

Improvements

• Add better hyperparameter search (learning rate, buffer size, exploration decay)
• Add experience replay inspection and prioritized replay
• Utilize GPU with tensorboard logging and model checkpoints
• Train on full 49 Atari games with a shared/transfer learning agent