Reinforcement Learning for Atari

Experiments with Reinforcement Learning to play Atari games. See the original Google Deepmind Atari paper.

This repo is a modified fork of Reinforcement Learning repo authored by @BaptisteBell and @bribridu59.

Modifications & Improvements

The following list describes the diff between this fork and the original repo

Training:

• Double DQN
• Removed batch normalization
• Modified hyperparameters

Performance:

• PyTorch compilation
• Fused AdamW optimizer
• TF32 precision
• Gradient clipping

Quality of life:

• CSV exports for rewards and losses during each training run
• Commandline args for profiling and rendering

Usage

# Training
python pong.py
python pong.py --enable-perf-logs

# Render Demonstration of model gameplay
python pong.py --demo
python pong.py --demo -m ./models/good_model.pth

Results

Rendering

Green = good_model.pth

Loss @ 500 episodes (each plot is a unique training run) Reward @ 500 episodes (each plot is a unique training run)