Эксперимент: Robust Lyapunov Training Under Model Error (#1109)

{ "worked": false, "confidence": 9, "verdict": "Built a nonlinear 2D stabilization MVP with nominal and robust Lyapunov penalties, analytic local Lipschitz inflation, bounded model mismatch, mechanism sweeps, and matched data-scarcity comparisons. The Lipschitz inequality passed numerically, with maximum violation about -1.5e-7, and the delta sweep matched the analytic certificate degradation within about 2e-7. However, convergence remained near 1% and robust training did not improve it, so the promised practical effect was not demonstrated.", "metrics": { "baseline": "Convergence 0.0113, 0.0147, 0.0087 at data fractions 1.0, 0.5, 0.2; actual violation 0.549, 0.548, 0.535; minimum robust margins -0.623, -0.779, -0.944.", "idea": "Convergence 0.0093, 0.0133, 0.0087 at data fractions 1.0, 0.5, 0.2; actual violation 0.543, 0.533, 0.507; minimum robust margins -0.618, -0.751, -0.915. The delta mechanism sweep matched prediction, but all certificates were negative." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "The optimizer was deliberately small and used a fixed quadratic Lyapunov function rather than jointly learning V. Uncertainty was prescribed instead of estimated from an ensemble or calibrated held-out residuals. The controller was not stabilizing, so a meaningful zero-margin convergence transition was not located, and no larger RL benchmark was tested." }