Snake AI

I built this snake game with two AIs: a pathfinder that plans a route to the food, and a neural network that learns by trial and error. You can watch them play, compare them in the arena or take over the controls.

Speed 1×

The Pathfinder is playing with its tuned settings.

This game

Playing: Pathfinder
Score: 0
Efficiency: –
Best this visit: 0
All-time: 0

Efficiency is the shortest route to each piece of food divided by the route the snake took. Best scores are saved for each player in this browser.

League

Use GPU

Picked when training starts.

Not training. Start training to add rounds on this device.

Rounds: –
Games played: –
Balance, score / efficiency: –
Champions crowned in round: –

Pathfinder vs DQN, round by round

Pathfinder
DQN

Champion settings

Pathfinder
Setting	Value

DQN
Setting	Value

How it works

Pathfinder

Every move, a breadth-first search finds the shortest path to the food. Before taking it, the snake simulates eating and checks it could still reach its own tail. If not, it follows its tail until a safe route opens. If it goes too long without eating, it tries a riskier route.

DQN

I first trained this network in PyTorch on a Raspberry Pi 5 in January 2025. My first version, a plain DQN, climbed to a best score of 37 and then forgot what it had learned, so I rebuilt it as a Double DQN with a bigger replay memory and steadier exploration. That version, ported to JavaScript, is where the league starts.

The league

An evolution strategy tunes the Pathfinder's move weights. Population-based training adjusts the DQN settings, including learning rate and network size. Both play the same games, and a judge scores them on food collected and path efficiency and adjusts the weight of each measure. The DQN trains on Pathfinder moves, and its value estimates help guide the Pathfinder. Training runs in a background worker, using your graphics card when that's faster.