November 13th, 2006



heh. In experimental machine learning engineering, we do oracle experiments, which garner this evocative name because they are experiments in which we "cheat" and ask "suppose we had an oracle to make exactly the right decision. How well could we do?"

Unfortunately, my best experimental results are reporting an F-measure that is higher than my oracle experiments. This is a problem in one of two places: either my results from the experiment are artificially high (which would be sad) or my oracle computation is broken.

I thought I had fixed the oracle computation, but I am still seeing this problem.

Ah well.

I'm definitely enjoying the struggle, though.

And boobirdsfly is sailing along with NaPlWriMo. Go see what it's all about!