Hello David.
Here is the experiment I have currently tasked Juan Li with:
[mu is the location of the ARD prior's mean, and is a matrix]
1) mu=0, set random seed, run SSM, obtain ROC curve ROC-A.
2) choose an arc r->s which SSM doesn't get right at all.
3) plot ROC curve B, ROC-B, obtained from ROC-A but where arc r->s is
manually corrected.
4) rerun SSM (from same random seed as above), with mu_{r,s} =
[0 .1 .5 1 2 4 8 12 16]
5) in each case plot the new ROC curve C, {ROC-C}_{mu_{r,s}
=0 .1 .5 1 2 4 8 12 16}.
By construction, ROC-B must be at least as good as ROC-A.
Without local minima issues, we would hope that ROC-C were better
than ROC-A.
But we would hope that ideally ROC-C is better than ROC-B, which
would show that providing information about r->s provides *more than
just this isolated help*.
Will follow up on the discussion we had about the knock-out.
-Matt
