Local search-based online learning algorithm for shape and cross-section optimization of free-form single-layer reticulated shells

Abstract

Reasonable shape and cross-section design of free-form Single-Layer Reticulated Shells (SLRSs) are crucial for their superior static performance and material efficiency. However, traditional metaheuristics face high computational costs and are prone to converging to local optima when optimizing these factors simultaneously, often leading to necessity of carrying out decoupled design processes. This paper introduces a Local Search-based Online Learning Algorithm (LSOLA) for simultaneous shape and cross-section optimization of free-form SLRSs. LSOLA builds deep learning models in various sub-regions of the solution space and uses a hybrid query strategy to actively select promising samples, iteratively improving prediction accuracy near potentially optimal solutions for more efficient exploration. Numerical examples show that LSOLA delivers more diverse and superior solutions at lower computational costs compared to the existing global search-based online learning algorithms and metaheuristics. This paper also offers a reference for other optimization problems involving numerous variables and nonlinear constraints.