Knieper Rollout Stability Metric / bench_report.json

Failed on benchmark

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": 0.006,
 10      "lambda": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "lambda": 0.0
 17        },
 18        "mean": 0.5065934360027313
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "lambda": 0.0
 24        },
 25        "mean": 0.264720531180501
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "lambda": 0.0
 31        },
 32        "mean": 0.05216603260487318
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.04864278854802251,
 37      "std": 0.028737614707129207,
 38      "per_seed": [
 39        0.02175583690404892,
 40        0.0447821244597435,
 41        0.11459071189165115,
 42        0.02753545716404915,
 43        0.05169341713190079,
 44        0.06664472073316574,
 45        0.03892377018928528,
 46        0.02321626991033554
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.051210810197517276,
 53    "std": 0.027750063002216954,
 54    "per_seed": [
 55      0.02233128994703293,
 56      0.04393337294459343,
 57      0.1131841167807579,
 58      0.02899443544447422,
 59      0.06591296941041946,
 60      0.06486307084560394,
 61      0.038853198289871216,
 62      0.0316140279173851
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 0.002568021649494767,
 68    "idea_wins": 4,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      0.0005754530429840088,
 72      -0.0008487515151500702,
 73      -0.0014065951108932495,
 74      0.0014589782804250717,
 75      0.014219552278518677,
 76      -0.001781649887561798,
 77      -7.05718994140625e-05,
 78      0.00839775800704956
 79    ],
 80    "p_value": 0.3862,
 81    "mde": 0.00477516318712563,
 82    "mde_rel_pct": 9.816795725868714,
 83    "verdict": "no significant win",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "quantity": "median test-set G_H=max hidden separation / initial perturbation norm",
 88    "H": 8,
 89    "predicted": "rollout penalty reduces finite-horizon gain",
 90    "baseline_mean_GH": 0.5347721576690674,
 91    "idea_mean_GH": 0.5261865258216858,
 92    "observed_relative_reduction": 0.01605474728677747,
 93    "predicted_vs_observed": {
 94      "predicted_direction": "decrease",
 95      "observed_direction": "decrease"
 96    },
 97    "confirmed": false,
 98    "paired_seed_GH_baseline": [
 99      0.5360813736915588,
100      0.5294458866119385,
101      0.509493350982666,
102      0.5483631491661072,
103      0.5446451306343079,
104      0.5424107909202576,
105      0.5554588437080383,
106      0.5122787356376648
107    ],
108    "paired_seed_GH_idea": [
109      0.5199079513549805,
110      0.5318974256515503,
111      0.4956446886062622,
112      0.5407124161720276,
113      0.5468658208847046,
114      0.5303430557250977,
115      0.54232257604599,
116      0.5017982721328735
117    ],
118    "idea_config": {
119      "lr": 0.006,
120      "lambda": 0.01
121    },
122    "track_choice_justification": "dynamics is the built-in control/stability track with multi-step pendulum rollouts."
123  },
124  "idea_sweep": [
125    {
126      "cfg": {
127        "lr": 0.006,
128        "lambda": 0.01
129      },
130      "mean": 0.051210810197517276,
131      "per_seed": [
132        0.02233128994703293,
133        0.04393337294459343,
134        0.1131841167807579,
135        0.02899443544447422,
136        0.06591296941041946,
137        0.06486307084560394,
138        0.038853198289871216,
139        0.0316140279173851
140      ]
141    },
142    {
143      "cfg": {
144        "lr": 0.006,
145        "lambda": 0.03
146      },
147      "mean": 0.051318740006536245,
148      "per_seed": [
149        0.02230515517294407,
150        0.04414716735482216,
151        0.11361950635910034,
152        0.028880951926112175,
153        0.06606191396713257,
154        0.06464919447898865,
155        0.03913993015885353,
156        0.03174610063433647
157      ]
158    },
159    {
160      "cfg": {
161        "lr": 0.006,
162        "lambda": 0.1
163      },
164      "mean": 0.05190074094571173,
165      "per_seed": [
166        0.02226673997938633,
167        0.0442110151052475,
168        0.11737498641014099,
169        0.02898707240819931,
170        0.06658705323934555,
171        0.06442747265100479,
172        0.03898325189948082,
173        0.032368335872888565
174      ]
175    }
176  ],
177  "protocol_notes": {
178    "epochs": 5,
179    "n_train": 400,
180    "n_test": 200,
181    "baseline_lr_union": [
182      0.0015,
183      0.003,
184      0.006
185    ],
186    "idea_lambdas": [
187      0.01,
188      0.03,
189      0.1
190    ],
191    "cache_used": true
192  }
193}