Эксперимент: Resonance-Aware Stochastic RNN Control (#447)

{ "worked": true, "confidence": 8, "verdict": "Built a readable stochastic rotational hidden-state MVP with correlation-based damped-mode fitting, PSD peak estimation, finite-time Lyapunov estimation, parameter sweeps, and resonance-radius control. The mechanism manifested quantitatively: fitted decay-time error was at most 2.9%, spectral peak error at most 4.1%, and fitted resonance parameters were invariant to noise while variance changed substantially. Controlling radius from 0.975 to 0.90 reduced peak power by 14.9x and long-lag correlation by 12.7x, although this is an exactly analyzable toy system rather than a trained nonlinear RNN.", "metrics": { "baseline": "Uncontrolled r=0.975: fitted r=0.9753, decay time=39.93, PSD peak power=25.51, absolute lag-80 correlation=0.0623, Lyapunov=-0.0253.", "idea": "Controlled r=0.90: fitted r=0.8996, decay time=9.45, PSD peak power=1.71, absolute lag-80 correlation=0.00490, Lyapunov=-0.1054. Radius-sweep decay prediction maximum relative error=2.91%; frequency prediction maximum relative error=4.12%; noise-invariance spans were 0.00157 in fitted radius and 0.00284 rad in fitted frequency." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 resonance_experiment.py", "files": [ "resonance_experiment.py", "results/report.json", "results_stdout.json" ], "limitations": "The experiment uses a linear 2D noisy rotation, where the exact transfer resonance is known and Lyapunov estimation is straightforward; it does not train or compare nonlinear tanh/GRU RNNs on sine prediction or copy-memory tasks. The controller is implemented as direct recurrent-radius clipping rather than a differentiable periodically updated penalty, and no compute/FLOP or CUDA comparison was performed." }