Comparing Lamarckian and Baldwinian Approaches in Memetic Optimization

Jiaojiao Mei, László Gulyás, János Botzheim

Abstract: Memetic optimization (MO) unifies local and global search in non-monotonic, ’rugged’ search spaces. As an enhancement of genetic optimization, where global search is via crossover and local search through random mutation, MO employs advanced techniques for local search and its integration with the global counterpart. This paper compares two approaches to memetic optimization (MO) based on the Baldwinian and Lamarckian theories of evolution. These theories propose that individuals’ behaviors are enhanced not only through crossover and mutation but also via lifetime learning. The Baldwinian approach suggests that learned behaviors impact genotype-phenotype mappings, leading to changes in individuals’ fitness. In contrast, the Lamarckian approach posits that learned behaviors not only affect individual fitness but also transfer to offspring. Our study demonstrates that genetic algorithms outperform memetic algorithms in general unimodal and multimodal functions, whereas memetic algorithms utilizing the Baldwinian and Lamarckian approaches excel in fixed-dimension multimodal functions, uncovering more global minima and generating superior solutions. These findings, derived from extensive experiments on the CEC-BC-2017 test functions, provide valuable insights for algorithm selection and optimization in various problem domains.