JSI123/Reinforce-PixelCopter

REINFORCE Agent playing Pixelcopter-PLE-v0

A custom PyTorch implementation of the REINFORCE policy-gradient algorithm.

Training Details

• Environment: Pixelcopter-PLE-v0
• Algorithm: REINFORCE
• Training episodes: 63,000
• Policy architecture: 7 → 64 → 128 → 2
• Optimizer: Adam
• Learning rate: 0.00005
• Discount factor: 0.99
• Device: CPU

Evaluation Results

• Evaluation episodes: 10
• Mean reward: 43.90
• Standard deviation: 37.6761
• Leaderboard score: 6.2239
• Required score: 5.0

Files

• model.pt: Trained policy weights
• results.json: Evaluation results
• replay.mp4: Gameplay recording

Completed as part of Unit 4 of the Hugging Face Deep Reinforcement Learning Course.