Adaptive CBF Safety Layer for Neural Policies / results.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "math_check": {
  3    "min_true_minus_model_lower_bound": 9.999996386511611e-10,
  4    "statement": "true residual >= model residual when epsilon >= |d|"
  5  },
  6  "margin_threshold_sweep": [
  7    {
  8      "kappa": 0.8,
  9      "predicted_margin_crossing_step": null,
 10      "observed_crossing_step": null,
 11      "violations": 73
 12    },
 13    {
 14      "kappa": 1.0,
 15      "predicted_margin_crossing_step": null,
 16      "observed_crossing_step": null,
 17      "violations": 0
 18    },
 19    {
 20      "kappa": 1.2,
 21      "predicted_margin_crossing_step": 9,
 22      "observed_crossing_step": 9,
 23      "violations": 0
 24    },
 25    {
 26      "kappa": 1.5,
 27      "predicted_margin_crossing_step": 5,
 28      "observed_crossing_step": 5,
 29      "violations": 0
 30    },
 31    {
 32      "kappa": 2.0,
 33      "predicted_margin_crossing_step": 4,
 34      "observed_crossing_step": 4,
 35      "violations": 0
 36    },
 37    {
 38      "kappa": 3.0,
 39      "predicted_margin_crossing_step": 2,
 40      "observed_crossing_step": 2,
 41      "violations": 0
 42    }
 43  ],
 44  "smoothing_law_sweep": [
 45    {
 46      "rho": 0.5,
 47      "target_fraction": 0.8,
 48      "predicted_step": 3,
 49      "observed_step": 3
 50    },
 51    {
 52      "rho": 0.8,
 53      "target_fraction": 0.8,
 54      "predicted_step": 8,
 55      "observed_step": 8
 56    },
 57    {
 58      "rho": 0.9,
 59      "target_fraction": 0.8,
 60      "predicted_step": 16,
 61      "observed_step": 16
 62    },
 63    {
 64      "rho": 0.95,
 65      "target_fraction": 0.8,
 66      "predicted_step": 32,
 67      "observed_step": 32
 68    }
 69  ],
 70  "intervention_violation_sweep": [
 71    {
 72      "kappa": 0.0,
 73      "mean_violations": 87.875,
 74      "mean_intervention": 106.31445333387649,
 75      "mean_max_x": 1.3499635962177918
 76    },
 77    {
 78      "kappa": 0.5,
 79      "mean_violations": 83.0,
 80      "mean_intervention": 108.23935020540935,
 81      "mean_max_x": 1.1574739090645043
 82    },
 83    {
 84      "kappa": 1.0,
 85      "mean_violations": 0.0,
 86      "mean_intervention": 110.16424762161338,
 87      "mean_max_x": 0.9649841674441023
 88    },
 89    {
 90      "kappa": 1.5,
 91      "mean_violations": 0.0,
 92      "mean_intervention": 112.0891452641003,
 93      "mean_max_x": 0.7724944031954118
 94    },
 95    {
 96      "kappa": 2.0,
 97      "mean_violations": 0.0,
 98      "mean_intervention": 114.01404300310672,
 99      "mean_max_x": 0.5924474577665813
100    },
101    {
102      "kappa": 3.0,
103      "mean_violations": 0.0,
104      "mean_intervention": 117.86383848111953,
105      "mean_max_x": 0.40771009700058425
106    },
107    {
108      "kappa": 4.0,
109      "mean_violations": 0.0,
110      "mean_intervention": 121.71363395913237,
111      "mean_max_x": 0.317645356242228
112    }
113  ],
114  "baseline_static_clipping": {
115    "mean_violations": 92.0,
116    "mean_max_x": 11.981408929605442
117  },
118  "checks": {
119    "smoothing_matches_exact_recursion": true,
120    "intervention_monotone": true,
121    "violations_nonincreasing": true,
122    "adaptive_beats_clipping_at_kappa_2": true
123  }
124}