This talk starts by covering a variety of practical applications, with emphasis on those which some commercial enterprise actually wants to use for some reason rather than ``demonstrator" systems. There is a surprisingly wide variety of kinds of GA applications in use; and whereas the theory of genetic and other evolutionary algorithms very largely focusses on binary encodings and well-established generic algorithms, practical applications usually employ non-binary encodings and interesting hybrid algorithms, eg using local search rather than random mutation. That raises the question, why would you choose to use one if you can't prove much about it? This leads in turn to some discussion of how you might set about convincing yourself that it's going to work for you, and some observations about parameter sensitivity and other surprises. For example, we have a GA that works very well on large exam and lecture timetabling problems that involve non-binary as well as binary constraints, and it has been used in earnest for some years ... but it turns out that it can fail on certain very simple, solvable problems and yet work very reliably and fast on much larger examples from the same class. Should this worry you or not?

Back to Workshop Schedule

1996-1997 Mathematics in High Performance Computing