Your first game agent

Each decision is straight, turn left or turn right. The snake grows when it eats. Choose solid walls or wraparound edges in World. Hitting its own body always ends the round.

1. Open World and try the arrow keys or touch controls.
2. Choose Rule teacher to see a reference run.
3. Open Train to fit a small neural network without watching every attempt.
4. In Results, review the comparison, then choose Test in World to activate a passing model.

12 inputs → 24 neurons → 3 actions

Inputs: immediate danger in each direction, current heading, relative food position, and reachable free space after each action. For the full-board goal, the last three inputs are authored cycle-direction cues instead of free space. The cycle reference can fill these even-sized boards; the NN learns to follow that supplied route signal. Outputs: straight, left, right.

The neural network starts with random weights and imitates an authored rule teacher. No safety rule overrides its choice. Learning happens in Train; Run uses fixed weights.

Training and validation use different episodes. A checkpoint is selected using validation results before its final twelve test episodes. Success applies to this grid, selected wall rule and selected goal. Switching to pass-through edges requires its own training and checks.

Define the challenge, then train

Up to 6,000 examples, 60 epochs and a 60-second limit. Two candidate attempts per training round, including cancelled attempts.

Ready · 0 / 2 attempts used.

Activation needs at least 8/12 successes, at least two more than the untrained network, and no regression against a compatible active model. A failed candidate stays available for inspection.