On the behavioral diversity of random programs

Generating a random sampling of program trees with specified function and terminal sets is the initial step of many program evolution systems. I present a theoretical and experimental analysis of the expected distribution of uniformly sampled programs, guided by algorithmic information theory. This analysis demonstrates that increasing the sample size is often an inefficient means of increasing the overall diversity of program behaviors (outputs). A novel sampling scheme (semantic sampling) is proposed that exploits semantics to heuristically increase behavioral diversity. An important property of the scheme is that no calls of the problem-specific fitness function are required. Its effectiveness at increasing behavioral diversity is demonstrated empirically for Boolean formulae. Furthermore, it is found to lead to statistically significant improvements in performance for genetic programming on parity and multiplexer problems. Categories and Subject Descriptors I.2.2 [Artificial...

• Type: Conference
• Year: 2007
• Where: GECCO
• Authors: Moshe Looks
• Keywords: Artificial Intelligence, Behavioral Diversity, GECCO 2007, Program Evolution Systems