Size Control with Maximum Homologous Crossover

Michael Defoin Platel (1) , Manuel Clergue (1) , Philippe Collard (1)

1 I3S - Laboratoire d'Informatique, Signaux, et Systèmes de Sophia Antipolis

Most of the Evolutionary Algorithms handling variable-sized structures, like Genetic Programming, tend to produce too long solutions and the recombination operator used is often considered to be partly responsible of this phenomenon, called bloat. The Maximum Homologous Crossover (MHC) preserves similar structures from parents by aligning them according to their homology. This operator has already demonstrated interesting abilities in bloat reduction but also some weaknesses in the exploration of the size of programs during evolution. In this paper, we show that MHC do not induce any specific biases in the distribution of sizes, allowing size control during evolution. Two different methods for size control based on MHC are presented and tested on a symbolic regression problem. Results show that an accurate control of the size is possible while improving performances of MHC.

Domaines: Intelligence artificielle [cs.AI]

Michael Defoin Platel, Manuel Clergue, Philippe Collard. Size Control with Maximum Homologous Crossover. 7th International Conference, Evolution Artificielle, EA 2005, 2006, Lille, France. pp.13-24, ⟨10.1007/11740698⟩. ⟨hal-00159738⟩