Fly Tic-Tac-Toe
98-cell MaleCNS circuit + a 425-parameter readout in the browser, with silence / rewire controls. Imitation of minimax uses the circuit; a degree-preserving rewired graph scores about the same — topology is not the skill.
Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.
Dispute this rating →The evidence
No evidence, no level.The circuit update is a leaky tanh recurrence with stated constants (leak 0.3, gain 1.4, 3 iterations) - documented simplified dynamics, which is the L1 bar.
https://github.com/pikabell/fly-tictactoe/blob/bc5d32f971b28f6737059a837f4417ee3123d3cc/src/flybrain/dynamics.ts#L43-L62The only trainable object is a 425-parameter readout; the source states that no biological synapse changes anywhere.
https://github.com/pikabell/fly-tictactoe/blob/bc5d32f971b28f6737059a837f4417ee3123d3cc/src/flybrain/readout.ts#L1-L14The README says the graph, signs, dynamics and encoder are frozen and there is no plasticity anywhere in the model.
https://github.com/pikabell/fly-tictactoe/blob/bc5d32f971b28f6737059a837f4417ee3123d3cc/README.md#L29-L30The authors' own control: a rewired graph at equal training budget scores 0.850 vs the measured circuit's 0.854.
https://github.com/pikabell/fly-tictactoe/blob/bc5d32f971b28f6737059a837f4417ee3123d3cc/README.md#L48
Activity in the measured 98-cell MaleCNS circuit evolves under a documented simplified unit model - h <- 0.3*h + 0.7*tanh(u + 1.4*W^T h), three synchronous iterations, with the four constants named in one place - which clears L1. The README is explicit that h is dimensionless, neither a firing rate nor a membrane voltage. Not L2: the graph, the sign policy, the encoder and the dynamics are all frozen; the only thing any training touches is a 425-parameter readout (16 descending activities -> 16 hidden -> 9 squares), fitted either by minimax imitation or CEM, so no synapse inside the fly circuit is ever written. Worth carrying over: the project's own control table reports that a degree-preserving rewired graph, retrained at equal budget, scores 0.850 against the real circuit's 0.854 - i.e. the authors ran the topology control most projects skip, and it came out negative.
Record
Nearby
All projects →MaleCNS simulation connected to ViZDoom through modeled visual inputs and a fixed button readout. Includes plasticity experiments and negative validation results; learned survival has not been demonstrated.
Research framework comparing a MaleCNS-constrained sparse recurrent controller with rewired graphs, conventional neural networks, and other controls on simple VizDoom tasks.
Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.