Can Bayesian networks enhance genetic algorithms? This paper introduces the Bayesian Optimization Algorithm (BOA), a novel computational method designed to generate new candidate solutions by estimating the joint distribution of promising solutions. Positioned within the realms of genetic and evolutionary computation, the BOA employs Bayesian networks to effectively model multivariate data. Unlike traditional algorithms, the BOA adeptly identifies, reproduces, and mixes building blocks, irrespective of variable ordering. This method also enables the incorporation of prior problem information, though it's not a necessity. Experiments conducted on additively decomposable problems, with both overlapping and non-overlapping building blocks, showcased the BOA's ability to solve most problems in linear or near-linear time relative to problem size. This advancement marks a significant step toward addressing the challenge of correctly identifying and combining building blocks for complex problems with limited domain knowledge. The BOA's innovative approach has implications for optimization problems across various fields. It could also obtain good solutions for problems with very limited domain information.
As a contribution to Evolutionary Computation, this paper aligns with the journal's focus on innovative algorithms inspired by natural evolutionary processes. The use of Bayesian networks to enhance genetic algorithms directly addresses the journal's interest in improving computational problem-solving through evolutionary methods. The BOA algorithm promises increased efficiency and accuracy in optimization, resonating with the core themes of the journal.