Size-based tournaments for node selection

In genetic programming, the reproductive operators of crossover and mutation both require the selection of nodes from the reproducing individuals. Both unbiased random selection and Koza 90/10 mechanisms remain popular, despite their arbitrary natures and a lack of evidence for their effectiveness. It is generally considered problematic to select from all nodes with a uniform distribution, since this causes terminal nodes to be selected most of the time. This can limit the complexity of program fragments that can be exchanged in crossover, and it may also lead to code bloat when leaf nodes are replaced with larger new subtrees during mutation. We present a new node selection method that selects nodes based on a tournament, from which the largest participating subtree is selected. We show this method of size-based tournaments improves performance on three standard test problems with no increases in code bloat as compared to unbiased and Koza 90/10 selection methods. Categories and Subj...

• Type: Journal
• Year: 2011
• Where: GECCO
• Authors: Thomas Helmuth, Lee Spector, Brian Martin
• Keywords: Automatic Programming, GECCO 2011, Heuristic Methods, Optimization, Program Fragments