LIVE ARENA
The classic circuit SCATTER
SCORE00000
PELLETS LEFT—
POWER—
STEP000
Preparing the neural arcade…
Loading the latest experiment…
Arrow keys / WASD to move · Space to pause · one life · 2,400-step limit · side tunnel wraps
EXPERIMENT 03 · LEARNING THROUGH CONSEQUENCES
Pellets are rewards. Ghosts are lessons.
A fly-connectome controller enters the arcade.
THE SCORECARD · MEASURED, NOT SCRIPTED
The fly wiring is fixed. A motor readout learns from rewards, with no expert action labels. Food direction is an engineered graph-distance cue; this is assisted navigation, not raw visual perception.
Final evaluation uses separate ghost seeds on this arena, not unseen layouts.
Measured FlyWire connectivity drives a custom spiking-neuron model. For speed, training and live play reuse a bank of recorded neural responses to engineered sensory patterns. A numerical feature transform stabilizes reward learning. Only the artificial motor readout is trained; anatomical synapses remain unchanged. The reward signal is inspired by dopamine's teaching role, but no biological dopamine circuit is modeled. These results apply to this classic-layout simulation, not arcade hardware emulation or a living fly. Movement timing is scaled to decision steps. The original upward-target overflow quirks and level progression are not emulated. The character graphics visualize actual game states.
Ghost behavior reference: The Pac-Man Dossier by Jamey Pittman.