Looking for feedback on my GPU-accelerated Snake AI project [P]

Project Overview

I've been building an AI that learns to play the classic Snake game through reinforcement learning. The goal is to reach high scores while keeping training time as low as possible.

The current version averages 86 points (87 is the maximum) after less than 10 hours of training on a single free Google Colab T4 GPU.

Training Approach

To keep training fast, it runs 4,096 Snake games directly on the GPU, combines GPU-native environment simulation with PPO + GAE, and uses a spatially-preserving CoordConv architecture that maintains the full game grid throughout training.

Request for Feedback

I'm sure there's still room to improve. If you've worked on reinforcement learning or efficient training systems, what would you try next? Better exploration, reward design, network architecture, or something else?

Repository: ( https://github.com/siddhartha399/PPO-CoordConv-Snake )

I'd really appreciate any feedback or criticism.