Inverse-Gain Structured Privileged Distillation / verification_report.json
Beats tuned baseline
1{
2 "seed": 2086,
3 "mechanism": {
4 "identity_max_abs_error": 4.263256414560601e-14,
5 "factor_identity_max_abs_error": 4.797551245161458e-14,
6 "disturbance_sweep": [
7 {"std": 0.0, "max_error": 3.552713678800501e-14},
8 {"std": 0.1, "max_error": 4.263256414560601e-14},
9 {"std": 0.5, "max_error": 4.973799150320701e-14},
10 {"std": 1.0, "max_error": 4.263256414560601e-14},
11 {"std": 2.0, "max_error": 4.973799150320701e-14},
12 {"std": 4.0, "max_error": 4.263256414560601e-14}
13 ],
14 "inverse_gain_noise_sweep": {
15 "mean_abs_z": 4.352920169020208,
16 "rows": [
17 {"q_error_std": 0.01, "observed_mae": 0.03486036026051096, "predicted_mae": 0.034782},
18 {"q_error_std": 0.03, "observed_mae": 0.10369104797490225, "predicted_mae": 0.104346},
19 {"q_error_std": 0.06, "observed_mae": 0.20862916908206103, "predicted_mae": 0.208692},
20 {"q_error_std": 0.12, "observed_mae": 0.4154603849890187, "predicted_mae": 0.417384},
21 {"q_error_std": 0.24, "observed_mae": 0.8331213634687362, "predicted_mae": 0.834768}
22 ],
23 "note": "Predicted MAE uses std*E|z|*sqrt(2/pi) for Gaussian q error."
24 }
25 },
26 "offline_action_imitation": {
27 "mlp_rmse_mae": [0.5011866092681885, 0.3038865327835083],
28 "gru_rmse_mae": [0.5548794865608215, 0.3636118769645691],
29 "structured_gru_rmse_mae": [0.5133934617042542, 0.31270310282707214]
30 }
31}