A comparison of an estimation of distribution algorithm and a stochastic hill-climber for composite optimization problems

Evolutionary algorithms (EA) have become a standard tool for the optimization of complex composite structures because of their ability to solve combinatorial problems. However, several studies have shown that simpler algorithms, such as stochastic hill climbers (SHC) can be more efficient even on problems designed to demonstrate EAs superiority, such as the Royal Road problem. The present paper compares the performance of a variant of EA, the univariate marginal distribution algorithm (UMDA) with that of an SHC on different fitness landscapes found in laminate optimization problems and identifies factors that influence the algorithms’ relative performance. In particular, it is found that mUMDA, a hybrid algorithm that combines UMDA’s global distribution learning and SHC’s local random search, outperforms SHC on large, highly constrained problems and on multimodal problems.