Koopman-MPC Trust Region for Neural Rollouts / report_bench_2026-09-03T132246.md
Mechanism confirmed, baseline not beaten
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": false, "confidence": 9, "verdict": "Implemented and ran the required 8-seed Stage-2 benchmark on the structurally matched dynamics track with identical rnn_small systems. The Koopman trust-region consistency penalty confirmed the stabilization mechanism but worsened test MSE, so there is no demonstrated task-level win.", "metrics": { "baseline": "Tuned rnn_small, Adam lr=0.01, 16 epochs: mean test MSE 2.410371e-05, std 9.965205e-06.", "idea": "Koopman spectral-cap/trust-region penalty, lr=0.01 and strength=0.1: mean test MSE 3.297570e-05, std 8.535059e-06; paired delta +8.871987e-06, permutation p=0.06005, 2/8 wins. Mechanism: raw fitted rho=0.993781, capped rho=0.95; NN deviation from the fitted stable forecast decreased from 0.035890 baseline to 0.031534 idea, confirmed=true." }, "bench_report": { "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01, "weight_decay": 0.0 }, "sweep": [ { "cfg": { "lr": 0.001, "weight_decay": 0.0 }, "mean": 0.0006648988055530936 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0 }, "mean": 0.00014707197260577232 }, { "cfg": { "lr": 0.01, "weight_decay": 0.0 }, "mean": 2.476238842064049e-05 } ], "full": { "mean": 2.4103710074996343e-05, "std": 9.965204787101785e-06, "per_seed": [ 4.82606228615623e-05, 1.8373453713138588e-05, 1.6678590327501297e-05, 1.573688678035978e-05, 1.862385761342011e-05, 2.8347380066406913e-05, 2.4800614482956007e-05, 2.2008274754625745e-05 ], "n": 8 } }, "idea": { "mean": 3.29756971950701e-05, "std": 8.53505931418874e-06, "per_seed": [ 3.491392271826044e-05, 3.567166277207434e-05, 1.528442589915357e-05, 3.08263479382731e-05, 3.304120764369145e-05, 4.200598777970299e-05, 4.480403367483718e-05, 2.7257989131612703e-05 ], "n": 8, "selected_cfg": { "lr": 0.01, "strength": 0.1 } }, "comparison": { "delta_mean": 8.871987120073754e-06, "idea_wins": 2, "n_pairs": 8, "per_seed_diffs": [ -1.334670014330186e-05, 1.729820905893575e-05, -1.3941644283477217e-06, 1.508946115791332e-05, 1.4417350030271336e-05, 1.365860771329607e-05, 2.000341919483617e-05, 5.249714376986958e-06 ], "p_value": 0.06005, "mde": 9.478008368122907e-06, "mde_rel_pct": 39.32178215981278, "verdict": "no significant win", "system_worked": false }, "mechanism_signature": { "baseline": { "prediction_abs": 0.005701124202460051, "koopman_abs": 0.03589034453034401, "rho_capped": 0.9499999284744263, "rho_raw": 0.993781371389095 }, "idea": { "prediction_abs": 0.004197320435196161, "koopman_abs": 0.03153369948267937, "rho_capped": 0.9499999284744263, "rho_raw": 0.993781371389095 }, "predicted_effect": "spectral cap keeps fitted transition rho <= 0.95 and lowers NN deviation from the fitted stable forecast", "confirmed": true }, "config": { "epochs": 16, "n_train": 2000, "n_test": 500, "baseline_grid": [ 0.001, 0.003, 0.01 ], "idea_grid": [ { "lr": 0.001, "strength": 0.02 }, { "lr": 0.003, "strength": 0.05 }, { "lr": 0.01, "strength": 0.1 } ] } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_bench.py", "files": [ "stage2_bench.py", "bench_report.json" ], "limitations": "Only the built-in actuated-pendulum dynamics track was tested; no custom track, longer horizons, alternative QP solver, explicit online receding-horizon control, wall-clock speed, or larger model was evaluated. The implementation uses a training-time Koopman consistency penalty and a clipped trust-region target rather than a full constrained MPC QP.", "system_verdict": "partial", "practical_verdict": "inconclusive", "mechanism_ok": 1, "system_judged": true }