Koopman-MPC Trust Region for Neural Rollouts / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "rnn_small",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01,
 10      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.0006648988055530936
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 0.00014707197260577232
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 2.476238842064049e-05
 33      }
 34    ],
 35    "full": {
 36      "mean": 2.4103710074996343e-05,
 37      "std": 9.965204787101785e-06,
 38      "per_seed": [
 39        4.82606228615623e-05,
 40        1.8373453713138588e-05,
 41        1.6678590327501297e-05,
 42        1.5736886780359782e-05,
 43        1.8623857613420114e-05,
 44        2.8347380066406913e-05,
 45        2.4800614482956007e-05,
 46        2.2008274754625745e-05
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 3.29756971950701e-05,
 53    "std": 8.53505931418874e-06,
 54    "per_seed": [
 55      3.491392271826044e-05,
 56      3.567166277207434e-05,
 57      1.5284425899153575e-05,
 58      3.08263479382731e-05,
 59      3.304120764369145e-05,
 60      4.200598777970299e-05,
 61      4.4804033677792177e-05,
 62      2.7257989131612703e-05
 63    ],
 64    "n": 8,
 65    "selected_cfg": {
 66      "lr": 0.01,
 67      "strength": 0.1
 68    }
 69  },
 70  "comparison": {
 71    "delta_mean": 8.871987120073754e-06,
 72    "idea_wins": 2,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      -1.334670014330186e-05,
 76      1.7298209058935754e-05,
 77      -1.3941644283477217e-06,
 78      1.508946115791332e-05,
 79      1.4417350030271336e-05,
 80      1.3658607713296078e-05,
 81      2.000341919483617e-05,
 82      5.249714376986958e-06
 83    ],
 84    "p_value": 0.06005,
 85    "mde": 9.478008368122907e-06,
 86    "mde_rel_pct": 39.32178215981278,
 87    "verdict": "no significant win",
 88    "system_worked": false
 89  },
 90  "config": {
 91    "epochs": 16,
 92    "n_train": 2000,
 93    "n_test": 500,
 94    "idea_grid": [
 95      {
 96        "lr": 0.001,
 97        "strength": 0.02
 98      },
 99      {
100        "lr": 0.003,
101        "strength": 0.05
102      },
103      {
104        "lr": 0.01,
105        "strength": 0.1
106      }
107    ]
108  },
109  "mechanism_signature": {
110    "baseline": {
111      "prediction_abs": 0.005701124202460051,
112      "koopman_abs": 0.03589034453034401,
113      "rho_capped": 0.9499999284744263,
114      "rho_raw": 0.993781371389095
115    },
116    "idea": {
117      "prediction_abs": 0.004197320435196161,
118      "koopman_abs": 0.03153369948267937,
119      "rho_capped": 0.9499999284744263,
120      "rho_raw": 0.993781371389095
121    },
122    "predicted_effect": "spectral cap keeps fitted transition rho <= 0.95 and lowers NN deviation from the fitted stable forecast",
123    "confirmed": true
124  }
125}