Running live in your browser

A neural network that learns Snake.

No PyTorch, no black box. I wrote the forward pass, the loss and backprop by hand in NumPy, gave the network a game of Snake and nothing else. Pick a checkpoint and watch it get smarter.

shape: 11 → 64 → 3
method: Q-learning, replay, Adam
stack: NumPy only, hand-written backprop
hosting: static, $0

Fig. 1

Live game

Score0

Best0

Games0

Avg 100

Fig. 2

What the network sees

excitatoryinhibitoryfilled = firing now

Fig. 3

Training

score per game, 400 games, 10-game average

Best training game

Greedy avg · 30

Weights963

Notes

How it works

01Forward pass

11 yes/no inputs: danger ahead, right and left, heading, and where the food is. One hidden layer picks one of 3 moves.

h = max(0, x @ W1 + b1)
q = h @ W2 + b2
# action = argmax(q)

02Q-learning

Eating is +10, crashing is −10. The network learns to predict how good each move is, from its own past games.

y = r + 0.9 * max(Q(s'))
loss = (Q(s,a) - y)^2
# replay buffer, Adam

03Backprop

Every crash sends an error backwards through the layers. Each gradient is written by hand.

dq  = 2*(q - y)
dW2 = h.T @ dq
dh  = (dq @ W2.T) * (h > 0)
dW1 = x.T @ dh