Эксперимент: Behavior-Gap Clustered Neural Controllers (#1323)

{ "worked": false, "confidence": 9, "verdict": "Built a self-contained behavior-gap controller prototype with Hankel trajectory construction, rank-reduced behavior projectors, greedy clustering, and shared Adam moments. The projector-gap identity was verified to machine precision, and the toy systems separated into the intended two behavior groups at horizons 8 and 16, although horizon 4 was less reliable. Sharing optimizer moments increased final loss from 1.85e-11 for independent Adam to 9.82e-2, so the proposed mechanism did not produce a win in this controlled test.", "metrics": { "baseline": "Independent Adam: final MSE 1.85e-11, 220 steps, 8.13 s", "idea": "Behavior-gap clusters with shared Adam moments: final MSE 9.82e-2, 220 steps, 7.03 s; loss ratio 5.31e9x baseline", "math": "Projector-gap identity absolute error 0.0; one-dimensional angle test matched sin(theta) exactly; L=8 formed clusters [0,1,2,3] and [4,5,6,7]" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only a tiny synthetic 2D recurrent system was tested, with one fixed seed and short finite-horizon trajectories. No GRU, sequential-MNIST, CUDA benchmark, repeated-seed statistics, hysteresis, compatibility-margin filtering, low-rank adapters, or controller-sharing variant was evaluated." }