Inverse-Gain Structured Privileged Distillation / results.json
Beats tuned baseline
1{
2 "math_checks": {
3 "identity_max": 4.263256414560601e-14,
4 "factor_identity_max": 4.797551245161458e-14,
5 "sigmas": [
6 0.01,
7 0.03,
8 0.06,
9 0.12,
10 0.24
11 ],
12 "maes": [
13 0.03476888668038175,
14 0.1045646578860718,
15 0.20821952296751994,
16 0.41841178608666424,
17 0.8347931744420186
18 ],
19 "slope": 3.479225134076925,
20 "predicted_slope": 3.4806462926256576
21 },
22 "action_rmse_mae": {
23 "mlp": [
24 0.5011866092681885,
25 0.3038865327835083
26 ],
27 "gru": [
28 0.5548794865608215,
29 0.3636118769645691
30 ],
31 "structured": [
32 0.5133934617042542,
33 0.31270310282707214
34 ]
35 },
36 "notes": "Exact identities use the sampled plant equation; learning comparison is teacher-forced offline action prediction."
37}