Эксперимент: Position-only active-noise optimizer (#1235)

{ "worked": true, "confidence": 8, "verdict": "Built a compact position-only active-noise optimizer prototype using a scalar OU disturbance and Kalman filtering of observed gradient residuals. The numerical stability boundary matched the augmented transition-matrix prediction exactly, with eta crossing at 2.0 when rho=1.0. Across five seeds, cancellation reduced stationary parameter MSE by about 97% versus SGD for all tested persistence times, so the claimed colored-noise benefit is clearly present in this toy setting.", "metrics": { "baseline": "At eta=0.3, mean parameter MSE across seeds: tau=0.2 0.1855, tau=1 0.3058, tau=5 0.6532, tau=20 0.8664.", "idea": "At eta=0.3 and cancellation gain c=0.85, mean MSE: tau=0.2 0.00618, tau=1 0.00950, tau=5 0.01916, tau=20 0.02517; these are respectively 3.33%, 3.11%, 2.93%, and 2.90% of baseline. Stability crossing: eta=2.000, rho=1.000." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 active_noise_optimizer.py && /home/maxwelhelp/main/bin/python3 run_repro.py", "files": [ "active_noise_optimizer.py", "run_repro.py", "toy_results.json", "toy_results_v2.json", "repro_summary.json" ], "limitations": "Only a one-dimensional quadratic system was tested. Curvature, OU persistence, disturbance variance, and measurement noise were supplied rather than estimated; the Kalman model was correctly specified, and the controller gain was manually chosen. No MLP/MNIST experiment, work-like objective measurement, multidimensional parameter-block filter, runtime comparison, or robustness test under model mismatch was performed." }