Snake AI

The idea

Nobody tells the agent how to play Snake. It only sees the game state and a reward: positive for eating, negative for crashing. Everything else, from chasing food to avoiding its own tail, has to be learned.

How it works

• Game. A classic Snake implementation in Python using curses, with movement, collision detection and scoring.
• Environment. A custom OpenAI Gym environment wraps the game: it defines the state representation, the action space (up, down, left, right) and the reward function.
• Agent. A Deep Q-Network built with TensorFlow estimates the value of each action. It explores with an ε-greedy policy that decays from 1.0 toward 0.01, and learns from mini-batches sampled out of a 2,000-step replay memory, which stabilizes training.

Training runs for 1,000 episodes and checkpoints the model every 50. After training, the snake plays on its own using the learned policy.

Why it matters

Reward design is the hard part: a badly shaped reward teaches the snake to circle forever instead of eating. Debugging behaviour rather than code is a different discipline, and the same intuition applies to any learning system deployed in the real world.