Meta-Brain

Uses the male CNS connectome as a sparse recurrent substrate for reinforcement learning: game frames are encoded into sensory neurons, activity propagates through the connectome-derived network, and a policy-gradient-trained readout decodes motor activity into buttons. Two tracks are reported honestly. A 40,000-neuron, 1,199,959-edge sensorimotor subgraph on a corridor task improved from -2.08 to 7.13 reward, success rate 25%. Full-connectome Super Mario reached about x=1,800 in some training episodes but averaged x=314 in evaluation, success rate 0, all five episodes ending in death.

Official site ↗

#reinforcement-learning#mario#negative-result#cuda#policy-gradient#malecns