This project is an implementation of the NEAT (NeuroEvolution of Augmenting Topologies) algorithm to train a neural network to play the classic Snake Game.

git clone https://github.com/your-username/neat-snake-ai.git
cd neat-snake-ai

# For Version 1
python SnakeV1.py

# For Version 2
python SnakeV2.py

You'll be prompted to enter a mode:

> Type 'train' to train a new model
> Type 'demo' to run the pretrained best genome

The AI sees the world in a minimal format and reacts with simple choices:

• Danger straight
• Danger left
• Danger right
• Food straight
• Food left
• Food right
• (+ Last output → only in Version 2)

• [1, 0, 0] → Go straight
• [0, 1, 0] → Turn left
• [0, 0, 1] → Turn right

No memory of the past, just pure reaction.

• -10 for dying
• +0.1 per time step
• +10 per apple
• Max Score: 98
• 50+ Score Achieved at Gen: 168

• -100 for dying
• Rewards unchanged
• Max Score: 94
• 50+ Score Achieved at Gen: 92

• +100 per apple
• +0.1 per time step
• Final fitness: (score^2) × 10
• Max Score: 52
• 50+ Score Achieved at Gen: 311

• +10 per apple
• +0.1 per time step
• No penalty for death
• Reward/Penalty for change in distance to food: ±0.5 × Δdist
• Max Score: 65
• 50+ Score Achieved at Gen: 337

The snake remembers its last move. Just enough memory to be dangerous.

• -10 for dying
• +0.1 per time step
• +10 per apple
• Max Score: 67
• 50+ Score Achieved at Gen: 237

• +10 per apple
• +0.1 per time step
• No penalty for death
• Reward/Penalty based on ±0.5 × Δdist
• Max Score: 74
• 50+ Score Achieved at Gen: 384

Each run tracks:

• Generation
• Best Fitness
• Max Apples Eaten
• Fitness History (Plotted with matplotlib)

The training results vary heavily depending on fitness function design — a small tweak can make or break evolution.

• Python
• Pygame – For rendering the snake game
• NEAT-Python – Evolutionary training
• Matplotlib – Plotting training progress
• Pickle – Saving top-performing genomes

• NEAT can evolve strong strategies even with minimal inputs.
• Reward shaping is crucial — a good scoring rule accelerates evolution.
• Adding simple memory (like last move) gives big gains in behavior consistency.

Evolution isn't perfect. But give it enough generations and even a snake becomes a strategist.