Play Mario with Reinforcement Learning

A Double Deep Q-Learning implementation that teaches an AI agent to play Super Mario Bros through trial and error. I wrote a blog post about this project here!

View on GitHub

Built With

This project implements a Double Deep Q-Learning (DDQN) model to teach an AI agent to play Super Mario Bros. The agent learns through trial and error, developing strategies to navigate levels, avoid enemies, and maximize score. The implementation uses PyTorch for deep learning and OpenAI Gym for the game environment.

The project involved several key components:

Environment Setup:

Integration with OpenAI Gym's Super Mario Bros environment
State preprocessing and action space definition
Reward function design to encourage desired behaviors

Model Architecture:

Implementation of a DDQN model with target network for stability
Experience replay buffer to store and sample past experiences
Training loop to update the model based on rewards and penalties

Training Pipeline:

Training the agent on episodes of Super Mario Bros with Epsilon-greedy exploration strategy
Monitoring performance metrics and checkpoints
Periodic target network updates