Uniform Stochastic Barrier Critic / results.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "seed": 11,
  3  "system": {
  4    "mu": 0.055,
  5    "sigma": 0.2,
  6    "U": -0.8,
  7    "T": 0.8
  8  },
  9  "predictions": {
 10    "margin_slope_observed": 15.958429160401632,
 11    "margin_prediction": "positive approximately linear until infeasible",
 12    "mc_slope_observed": -0.5113969907459289,
 13    "mc_prediction": -0.5,
 14    "feasibility_cliff_between_delta": [
 15      0.03,
 16      0.04
 17    ]
 18  },
 19  "margin_sweep": [
 20    {
 21      "delta": 0,
 22      "min_initial_B": 0.1358980998171118,
 23      "p_cert": 0.8641019001828882,
 24      "feasible": true
 25    },
 26    {
 27      "delta": 0.002,
 28      "min_initial_B": 0.16781495813790323,
 29      "p_cert": 0.8321850418620967,
 30      "feasible": true
 31    },
 32    {
 33      "delta": 0.005,
 34      "min_initial_B": 0.21569024561910813,
 35      "p_cert": 0.7843097543808919,
 36      "feasible": true
 37    },
 38    {
 39      "delta": 0.01,
 40      "min_initial_B": 0.2954823914212426,
 41      "p_cert": 0.7045176085787574,
 42      "feasible": true
 43    },
 44    {
 45      "delta": 0.02,
 46      "min_initial_B": 0.455066683025447,
 47      "p_cert": 0.5449333169745529,
 48      "feasible": true
 49    },
 50    {
 51      "delta": 0.03,
 52      "min_initial_B": 0.6146509746296235,
 53      "p_cert": 0.3853490253703765,
 54      "feasible": true
 55    },
 56    {
 57      "delta": 0.04,
 58      "min_initial_B": null,
 59      "p_cert": null,
 60      "feasible": false
 61    },
 62    {
 63      "delta": 0.05,
 64      "min_initial_B": null,
 65      "p_cert": null,
 66      "feasible": false
 67    }
 68  ],
 69  "mc_true_expectation": 0.5774997243675102,
 70  "mc_scaling": [
 71    {
 72      "M": 8,
 73      "mae": 0.02787765295299081
 74    },
 75    {
 76      "M": 16,
 77      "mae": 0.020605521927450753
 78    },
 79    {
 80      "M": 32,
 81      "mae": 0.014353076400914308
 82    },
 83    {
 84      "M": 64,
 85      "mae": 0.010239305597520103
 86    },
 87    {
 88      "M": 128,
 89      "mae": 0.007574911191649707
 90    },
 91    {
 92      "M": 256,
 93      "mae": 0.004955760000340562
 94    },
 95    {
 96      "M": 512,
 97      "mae": 0.0035631615262848772
 98    },
 99    {
100      "M": 1024,
101      "mae": 0.0022852087134715817
102    }
103  ],
104  "exact_failure_max_X0": 0.12995081649276266,
105  "rollout_success_by_start": [
106    0.8643864386438644,
107    0.8907890789078908,
108    0.8955895589558955,
109    0.9102910291029103,
110    0.9333933393339334,
111    0.9378937893789379,
112    0.9450945094509451,
113    0.9543954395439544,
114    0.9567956795679567
115  ],
116  "certificate_check": {
117    "delta": 0.03,
118    "max_B_X0": 0.6146509746296235,
119    "max_exact_failure_X0": 0.12995081649276266,
120    "conservative_gap": 0.48470015813686085
121  },
122  "neural_comparison": {
123    "device": "cuda",
124    "baseline": {
125      "loss": 0.3060360252857208,
126      "max_init_B": 0.22222056984901428,
127      "max_init_limit_violation": -0.7577794790267944,
128      "max_drift_residual": 0.28713932633399963,
129      "boundary_rmse": 0.1437969207763672
130    },
131    "idea": {
132      "loss": 8.059223175048828,
133      "max_init_B": 0.009476847015321255,
134      "max_init_limit_violation": -0.9705231785774231,
135      "max_drift_residual": 0.2172824889421463,
136      "boundary_rmse": 0.00011445609561633319
137    }
138  }
139}