Ask the Algorithm – Math with Bad Drawings

I am! It’s a really fun result (“train” yourself on the first 1/e of the talent pool, then select the first candidate who is better than all your training data). But I’ve always been somewhat troubled by the choice of optimization criterion.

It seems strange to maximize the probability of the single best candidate, as opposed to (say) expected quality of candidate. As such, the “optimal” algorithm has the peculiar feature that, 1/e of the time (when the best candidate happens to appear during your training period), you’ll wind up failing your search and selecting whichever candidate happens by random chance to be the final one interviewed.

In practice, a better approach seems to be relaxing your choosiness as you approach the end of the talent pool. E.g., you should accept a 2nd-to-last candidate if they’re merely better than the median candidate seen thus far (and accept a 3rd-to-last candidate if they’re in the top tercile of candidates seen thus far).