01 / PLATFORMER
Mario-style jumps
Can a network learn when to run and when to jump? Coins reward progress. Blocks, a gap, and an enemy make timing matter.
Play the platformer lab ↗LAB 02 / REINFORCEMENT LEARNING
Reinforcement
learning games
The games show how rewards affect an agent’s choices, alongside the neural network values used for each move.
The agent learns by trying moves and using feedback. You can change the settings while it plays, inspect real neuron activations, and compare a new brain with a practiced one. Everything runs in your browser. No account needed.
02 / MAZE
Pac-Man-style maze
The agent chooses moves in a maze with pellets, a chasing ghost, and power pellets that change the rewards.
Play the maze lab ↗A DIFFERENT WAY TO LEARN
Learning from rewards
In the movie-review lab, examples came with labels such as “positive” and “negative.” Here, no one supplies the correct move. The agent tries an action, receives a reward, and adjusts its guesses about which actions will pay off later. That is reinforcement learning.
Watch a fresh brain
Both games start playing immediately with random weights. Expect mistakes and lots of attempts.
Change one dial
Speed up practice, change random exploration, or adjust how much future rewards matter.
Inspect the neurons
The lights are actual values computed by the network for the move on screen. Pause and inspect one.
Compare with practice
Load a brain trained with the same algorithm. Watch its choices, then keep training it yourself.
These are small original browser games inspired by Mario and Pac-Man. The networks read game measurements, rather than pixels. Learning, games, and visuals run on your device.