Snake AI

Trained with Proximal Policy Optimization

0 Current Score

1 Episode

0 Food Eaten

This agent learned to play Snake through 37.1 million timesteps of reinforcement learning on a 20×20 grid. The model achieves a best score of 71 cells (17.8% grid fill) and an average of 32 cells across continuous play.

Training was conducted directly on the target grid size without curriculum learning, using 8 parallel environments and PPO with entropy regularization to encourage exploration.

Best Performance

71 cells

Training Steps

37.1M

Grid Fill

17.8%

Architecture

MLP

Architecture

Algorithm: Proximal Policy Optimization (PPO)
State space: 24 features (direction, food location, danger detection, connectivity)
Network: 2-layer MLP (64→64→4) with 6,020 parameters
Training: 8 parallel environments, batch size 256, learning rate 1e-4
Model size: 24KB (TensorFlow.js)

The agent demonstrates learned strategies including food-seeking behavior, obstacle avoidance, and basic space management. Performance continues to improve with extended training, approaching the theoretical ceiling for feature-based representations.