Neural Maze — Train an AI model to solve a maze

hyperparameters

personality presets

learning rate (α)0.30

How hard each new experience overwrites old beliefs.

discount (γ)0.950

How much the agent cares about far-away rewards.

exploration (ε)0.30

Chance of a random move. Follows the decay while training.

epsilon decay0.995

ε is multiplied by this after every episode.

maze size

layout

maze seed

Same seed + size = same maze, so you can compare settings fairly.

episodes per run

episode0

steps this ep0

reward0

success (last 20)0%

policy score0%

ε now0.30

cells known0

wall bumps0

Reach the goalnot yet

90% efficient policynot yet

Find the shortest routenot yet

shortest route: 0 steps

speednormal

keys: Space train/pause · S step · R reset · 1–5 speed · hover or tap any cell to inspect it

explore (random) exploit (learned) wall bump learned route lowhigh value

cell inspector

Hover or tap a cell to see what the agent believes about each move.