Snake Game AI

Reinforcement Learning (RL) is how the snake's AI learns — entirely through trial and error, with no human teacher. The agent (snake) takes an action, receives a numerical reward signal from the environment, and gradually figures out which actions lead to more total reward.

Think of it like training a dog: you don't explain the rules in words, you just give treats for good behaviour and none for bad. Over thousands of games the snake becomes competent.

Every action the snake takes returns a reward number. The AI's goal is to maximise the sum of future rewards:

The small per-step penalty discourages the snake from wandering aimlessly instead of seeking food.

The network receives a snapshot of 12 boolean (yes/no) features:

• Food direction — is food left / right / up / down of the head?
• Current heading — which way is the snake moving?
• Danger — is there a wall close by in each direction?

It outputs one Q-value per direction — a score estimating how much future reward that direction will yield. The highest-scoring safe direction wins.

DQN (Deep Q-Network) is the algorithm used in training mode. It faces a key trade-off:

• Explore — try random moves to discover new strategies
• Exploit — use what it already knows to score points

This is controlled by epsilon (ε). In training mode the network warm-starts with the pre-trained weights, so ε begins at 0.15 (mostly exploiting). It slowly decays toward 0.01 as fine-tuning proceeds.

DQN also stores past game transitions in an experience replay buffer and trains on random mini-batches. This breaks correlations between consecutive steps and stabilises learning.

Instead of starting from a blank neural network, the in-browser DQN is warm-started from the pre-trained model's weights. The weight shapes are identical (both use 24→64→64→4 layers after one-hot encoding), so the weights transfer directly.

This means the snake already knows the basics from episode 1. Training then fine-tunes rather than teaching from scratch, so improvements appear much sooner. Watch the Avg Score metric for evidence of continued learning.