Watch activations, inspect choices, and compare the brains you’ve trained.
0 training episodes · 160 weights
What goes in? What comes out?
Inputs: normalized X and Y position, horizontal and vertical direction to the trigger or exit, and four flags telling whether north, east, south, and west are open. The memory head also receives the exact maze, cell, and trigger-held state.
Outputs: four learned action values, shown as probabilities for North, East, South, and West. The highest legal action is chosen. Known states use learned memory weights; unseen states use the 8 → 12 → 4 dense network. There is no search fallback during a run.
One episode: up to 6,000 trial moves per source maze. Guided mode additionally trains on expert transitions and fits the demonstrated action. Exploration-only mode gets behavior rewards, but no route demonstrations. High guided accuracy means the training route was learned; it does not imply general intelligence or success on unseen mazes.
Train a brain to see rewards, wall hits, exploration, and dead-end recovery statistics.
POLICY CONFIDENCE—
CHOSEN DIRECTION—
DECISIONS OBSERVED0
TRAINING ACCURACY—
Live activations
8 → 12 → 4
Eight sensors feed the 12-neuron dense network. On learned cells, the state-memory head supplies action values instead. Node intensity shows activation; lines are not causal explanations.
Action probabilities
Waiting for a decision
Confidence is the policy’s probability for the chosen action, not a guarantee of reaching the exit. Blocked actions are masked when moving.
Confidence over time
LAST 120 DECISIONS
Learning curve
ACCURACY / TRAINING EPISODE
Where the choices flow
DIRECTION COUNTS
A younger mind. A different outcome.
Checkpoints are saved automatically during training. Compare route-teacher accuracy on this maze; then run a brain to test its actual behavior.
Train a network to create your first checkpoints.
↳ No two journeys are the same. Generate. Experiment. Get a little better.BUILT FOR THE BEAUTY OF THE PROCESS ✳
YOUR EXPERIMENTS, KEPT SAFE
My lab.
Every maze, every breakthrough, every almost.
MANY MINDS. ONE WAY OUT.
The scoreboard.
Valid routes. Real experiments. See how your runner stacks up.
Score = 10,000 × optimal steps / actual steps × grid size / 31 ÷ (1 + elapsed seconds / 120). Time includes animation and pauses. Routes are verified by the server; runner labels describe the selected mode.
A LITTLE DIRECTION
Meet your maze.
Part game. Part experiment. Entirely your playground.
01 / GENERATE
Make the challenge yours.
Choose a grid from 15 × 15 to 51 × 51. Hard mazes twist through deep corridors; Expert picks the longest route from three candidates. Easy favors straighter corridors.
Triggers & gates adds an amber key and a locked edge. Visit the key before crossing the gate. Every generated maze is solvable, and each roll is saved to My lab. All twists default to off. Black holes end a run when entered; a safe route always exists. Multi-trip mode requires 2 or 3 visits to the exit, returning to start between them. Fog of war reveals cells around your position and your explored trail; planners still know the full map.
02 / SOLVE
Choose a different mind.
A* prioritizes distance plus cost. Breadth-first search spreads in waves. Dijkstra explores the cheapest states first (equivalent to BFS on this unweighted grid). All three find a shortest legal route.
Depth-first search explores a branch before backtracking. It may take a longer route. Complete searches track the trigger and multi-trip leg. Try greedy best-first, weighted A*, beam search, random walking, greedy walking, left/right wall followers, and a least-visited explorer. Weak methods may loop, exhaust their budget, or enter black holes. Their failures are part of the lesson. Use Manual to play with arrow keys, WASD, or the touch controls.
YOUR CODE / MAZESCRIPT
Small instructions. Big ideas.
Write one instruction per line: N, E, S, W move one cell. RIGHT and LEFT take one wall-following step, initially facing east. Walls block movement.
REPEAT 10000
RIGHT
END
Repeat blocks can nest up to 8 levels and expand to at most 100,000 operations. Comments start with #. Use FORWARD or BACK relative to your heading; TURN LEFT / TURN RIGHT only rotates. IF OPEN FRONT (or N/E/S/W/LEFT/RIGHT/BACK), IF UNVISITED direction, IF KEY, IF DEADEND, IF AT START, and IF AT EXIT choose a block; ELSE is optional, END closes it. RANDOM, GREEDY, and EXPLORE choose a move. THINK pauses for one instruction, MARK records your cell, STOP ends the program. PLAN BFS / A* / DFS / GREEDY invokes a planner from your current state. All programs have a 100,000-instruction budget. Load the 11 annotated examples in Your code, then change a rule and compare.
NEURAL / EXPERIMENT
Teach a tiny network.
A hybrid policy runs in your browser: an 8 → 12 → 4 dense network plus learned state-action memory. Inputs include position, target direction, legal moves, exact cell identity, and trigger state. Outputs are the four movement choices. Memory lets it represent complex maze routes that a tiny dense network cannot.
Dense training adjusts 160 weights using backpropagation. The memory head learns state-action values through reward updates; guided mode also fits expert actions. Tune rewards for legal moves, new cells, leaving dead ends, triggers, and the exit, plus penalties for walls and revisits. Exploration-only learning can still loop or fail. There is no hidden algorithm fallback. Choose this maze’s expert route or a mixed set of your last five mazes, then tune reward, penalty, and learning rate. Checkpoints automatically preserve younger and older brains, weights, curriculum, and learning history. The Neural Observatory visualizes activations, action probabilities, confidence over time, and a Sankey diagram of choices. Compare checkpoint accuracy on the current maze and run each brain to test its actual behavior.
03 / RESULTS
Make your next run better.
Completed paths are replayed by the server, including gate rules. Scores reward shorter paths and faster completion, scaled by grid size. Pauses and animations count toward elapsed time.
Publish as ANON or use up to four letters or numbers. Every result includes a generated solve thumbnail. Filter the board by runner for a more useful comparison.
STAY A WHILE
Your own corner of the lab.
Guest profiles save your work for 30 days on this browser. Create an account to return from another device; registering keeps your guest experiments. Logging into another account switches libraries.
Tap the music note for a procedural ambient drone, soft noise, and playful solver sounds. Audio starts only after a tap. Fullscreen and touch movement are available on supported tablets.