Clustering-enhanced Bayesian optimization (CEBO): A case study on T-pedal wheel design of a stair-climbing robot

초록

This study proposes a clustering-enhanced Bayesian optimization (CEBO) framework to improve the efficiency and stability of Bayesian optimization (BO). The key idea is to construct a clustering-based trust region (CBTR) from the aggregated surrogate posterior means of multiple screened candidate points rather than relying solely on a single incumbent solution. This enables a data-driven and adaptive search process. In CEBO, candidate points are first generated across the design space and then screened adaptively using surrogate predictions. The screened candidates are then grouped by K-means clustering, with silhouette analysis used to determine the number of clusters. The CBTR is constructed from the selected cluster, and the next evaluation point is identified within this region by maximizing the expected improvement (EI) criterion. The screening ratio is adjusted according to the optimization progress, enabling focused local refinement while retaining a degree of exploration. Benchmark experiments show that CEBO achieves competitive and, in many cases, improved performance relative to established BO methods in terms of sampling efficiency, convergence reliability, and robustness. A case study on the T-pedal wheel design of a stair-climbing robot further illustrates its potential applicability in a realistic engineering optimization problem.

키워드