I first saw this suggested at http://donsbot.wordpress.com/2009/03/09/evolving-faster-hask... - he references prior work (see "best paper" in bold) by http://www.cis.udel.edu/~cavazos/index.php?page=publications

The ctuning site has a timeline that extends many years back, so I'm not suggesting it's "me too" material.

In (lossy) heuristic search, it's common to automatically tune a handful of parameters (with difficult to predict outcome that may depend on the specific data set) to find a good setting for a given valuation of time spent vs. search error.