nmaher's picture
Update logbook: Reproduction: STAND: Self-Aware Precondition Induction for Interactive Task Learning
558ca5f verified
|
Raw
History Blame Contribute Delete
1.45 kB

Claim 6: Hyperparameter sensitivity (lambda_p)


Verdict: TOY-VERIFIED

Script: lambda_p_sweep.py. 12 reps, hierarchical_shrinkage=True, λs=25/λn=50 fixed, λp swept over the paper's own values, accuracy + FP/FN reoccurrence at N=50 and N=100.

λp Accuracy FP reocc. FN reocc.
0 96.92% 6.65% 0.77%
0.5 96.99% 6.55% 0.79%
1.0 97.03% 6.55% 0.76%
5.0 98.28% (peak) 2.84% 0.16% (best)
10.0 97.84% 2.66% 0.71%
20.0 97.80% 2.66% 0.92%
50.0 97.26% 2.31% (best) 1.56%

Paper: "higher λp can benefit accuracy, productive monotonicity, and error reoccurrence. λp=25.0 retains these benefits while maintaining high precision near 100%." Direction matches cleanly: moving from λp=0 to any λp≥5 roughly halves FP reoccurrence and substantially improves accuracy, with a slight FN uptick only at the largest value tested (λp=50). This is consistent with the paper's own framing of λp=25 as a middle-ground choice rather than a monotonic "more is always better" relationship. λp=25 itself was not one of the paper's own swept values ({0, 0.5, 1, 5, 10, 20, 50}), so no direct single-point comparison is possible — matches the paper's own sweep grid.