Anyone who has used any algorithm with more than a couple of user-selectable parameters knows that “tuning” those parameters can be both important and NP-weird. One of the tutorials at this year’s meeting, “Meta-algorithms: From Algorithm Tuning and Configuration to Algorithm Portfolios” (presented by Meinolf Sellmann of IBM), demonstrated several computationally tractable approaches to automated tuning based on trials on test problem sets. The tutorial included both selection of “one size fits all” default values and problem-specific tuning (in which attributes of new problem instances are used to select one of a stable of predetermined configurations).
Performance gains, relative to default parameter settings, on SAT problems (using existing solvers) were quite impressive. Performance gains on your problems? Get the tutorial from the tutorials web site, have a look, and give it a try. A link for early access to the tutorials was emailed to all attendees. Members can access the tutorials through the society web site.