Эксперимент: Feasibility-Preserving Error Compensator (#1127)

{ "worked": true, "confidence": 8, "verdict": "Built and ran a deterministic constrained-servo simulator implementing the residual anti-windup compensator. Unsaturated decay matched the augmented spectral radius within 0.92% across the gain sweep, and the predicted stability boundary was 1.39 KI versus an observed finite-horizon boundary of 1.29. Saturation-duration scaling also matched the prediction: compensator growth slope fell from 0.00274 to 0.000414 per step, a 6.6x reduction, while constraint violation remained zero. In the fixed mini-test, post-release RMS error improved from 0.0969 for ordinary integral action to 0.0656 for the proposed method, although this remains a toy result rather than evidence of a general neural-network win.", "metrics": { "baseline": "No compensation: RMS error 0.5571, post-release RMS 0.0538, recovery 8 steps, max constraint violation 0.0", "idea": "Residual anti-windup: RMS error 0.5440, post-release RMS 0.0656, recovery 10 steps, max |c| 0.1604, max constraint violation 0.0; ordinary integral had post-release RMS 0.0969, recovery 19 steps, max |c| 0.7082" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 run_experiment.py", "files": [ "run_experiment.py", "results.json" ], "limitations": "No MLP, learned policy, multidimensional task space, noisy measurements, voltage/current/acceleration constraints, candidate shaper, or FLOP/runtime comparison was tested. The stability-boundary observation uses a finite-horizon decay criterion and differs from the exact spectral crossing because of threshold and grid effects." }