Projected Absolute-Residual Compensation for Neural State-Space Models / bench_report.json

✓✓ Beats tuned baseline

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      "gamma": 0.2
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "gamma": 0.2
 17        },
 18        "mean": 0.004167803970631212
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "gamma": 0.5
 24        },
 25        "mean": 0.004167803970631212
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.001,
 30          "gamma": 0.8
 31        },
 32        "mean": 0.004167803970631212
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "gamma": 0.2
 38        },
 39        "mean": 0.0035580022376962006
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.003,
 44          "gamma": 0.5
 45        },
 46        "mean": 0.0035580022376962006
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.003,
 51          "gamma": 0.8
 52        },
 53        "mean": 0.0035580022376962006
 54      },
 55      {
 56        "cfg": {
 57          "lr": 0.01,
 58          "gamma": 0.2
 59        },
 60        "mean": 0.0017142038705060259
 61      },
 62      {
 63        "cfg": {
 64          "lr": 0.01,
 65          "gamma": 0.5
 66        },
 67        "mean": 0.0017142038705060259
 68      },
 69      {
 70        "cfg": {
 71          "lr": 0.01,
 72          "gamma": 0.8
 73        },
 74        "mean": 0.0017142038705060259
 75      }
 76    ],
 77    "full": {
 78      "mean": 0.0015303929612855427,
 79      "std": 0.0006243748661918005,
 80      "per_seed": [
 81        0.0019577958155423403,
 82        0.0012857802212238312,
 83        0.0008496360969729722,
 84        0.00276360334828496,
 85        0.0014558577677235007,
 86        0.0009643874363973737,
 87        0.0009531342657282948,
 88        0.002012948738411069
 89      ],
 90      "n": 8
 91    }
 92  },
 93  "idea": {
 94    "mean": 4.955958911523339e-05,
 95    "std": 2.2175200246237042e-05,
 96    "per_seed": [
 97      2.1020921849412844e-05,
 98      4.073749732924625e-05,
 99      2.979358032462187e-05,
100      4.908094342681579e-05,
101      5.829468136653304e-05,
102      4.280764551367611e-05,
103      9.952809341484681e-05,
104      5.521334969671443e-05
105    ],
106    "n": 8
107  },
108  "comparison": {
109    "delta_mean": -0.0014808333721703093,
110    "idea_wins": 8,
111    "n_pairs": 8,
112    "per_seed_diffs": [
113      -0.0019367748936929274,
114      -0.001245042723894585,
115      -0.0008198425166483503,
116      -0.002714522404858144,
117      -0.0013975630863569677,
118      -0.0009215797908836976,
119      -0.000853606172313448,
120      -0.0019577353887143545
121    ],
122    "p_value": 0.0081,
123    "mde": 0.0005620586151503473,
124    "mde_rel_pct": 36.72642447846947,
125    "verdict": "idea better (significant)",
126    "system_worked": true
127  },
128  "mechanism_signature": {
129    "mechanism_signature": {
130      "n_models": 8,
131      "mean_abs_residual_norm": 0.7707589045166969,
132      "mean_incremental_residual_norm": 0.0819682078436017,
133      "mean_abs_inc_disagreement": 0.8208163678646088,
134      "predicted_relation": "incremental residual smaller than absolute residual under persistent mismatch",
135      "confirmed": true,
136      "note": "NN-scale behavior measured on trained dynamics models; built-in track has no explicit injected actuator-bias condition."
137    },
138    "idea_sweep": [
139      {
140        "cfg": {
141          "lr": 0.001,
142          "gamma": 0.2
143        },
144        "mean": 0.002574956219177693
145      },
146      {
147        "cfg": {
148          "lr": 0.001,
149          "gamma": 0.5
150        },
151        "mean": 0.0007634652483829996
152      },
153      {
154        "cfg": {
155          "lr": 0.001,
156          "gamma": 0.8
157        },
158        "mean": 9.187122088860633e-05
159      },
160      {
161        "cfg": {
162          "lr": 0.003,
163          "gamma": 0.2
164        },
165        "mean": 0.001222496688569663
166      },
167      {
168        "cfg": {
169          "lr": 0.003,
170          "gamma": 0.5
171        },
172        "mean": 0.00027883463826583466
173      },
174      {
175        "cfg": {
176          "lr": 0.003,
177          "gamma": 0.8
178        },
179        "mean": 8.01967735242215e-05
180      },
181      {
182        "cfg": {
183          "lr": 0.01,
184          "gamma": 0.2
185        },
186        "mean": 0.0005614861802314408
187      },
188      {
189        "cfg": {
190          "lr": 0.01,
191          "gamma": 0.5
192        },
193        "mean": 0.0001608945794941974
194      },
195      {
196        "cfg": {
197          "lr": 0.01,
198          "gamma": 0.8
199        },
200        "mean": 4.955958911523339e-05
201      }
202    ],
203    "structural_match": "actuated pendulum rollout/control and stability"
204  }
205}