next up previous
Next: Conclusions Up: Discussion Previous: Do we Want

Three Ways to Control Bloat

Three methods of controlling bloat have been suggested. Firstly, and most widely used (e.g. in these experiments), is to place a universal upper bound either on tree depth [16] or program length. ([23,24] discuss unexpected problems with this approach).

The second (also commonly used) is to incorporate program size directly into the fitness measure (often called parsimony pressure) [16,25,26]. [14] gives an analysis of the effect of parsimony pressure which varies linearly with program length. Multi-objective fitness measures where one objective is compact or fast programs have also been used [27].

The third method is to tailor the genetic operations. [28, page 469,] uses several mutation operators but adjusts their frequencies so a ``decrease in complexity is slightly more probable than an increase''. [10] suggests targeting genetic operations at redundant code. This is seldom used, perhaps due to the complexity of identifying redundant code. [7] showed bloat continuing despite their targeted genetic operations. Possibly this was because of the difficulty of reliably detecting introns. There was a route whereby the GP could evolve junk code which masqueraded as being useful and thereby protected itself from removal. While [14] propose a method where the likelihood of potentially disruptive genetic operations increases with parent size.



William B Langdon
Tue Jun 10 12:12:55 BST 1997