GA based construction of maximin latin hypercube designs for uncertainty design of experiment with dynamic strategy management

Abstract

Flexible construction of maximin Latin Hypercube Designs (LHDs) meets the NP-hard problem known as the Maximum Diversity Problem (MDP). Traditional algorithms, such as Genetic Algorithms (GAs), face challenges like premature convergence and limited optimization performance, particularly due to the number of hyperparameters that require to be tuned and their limited ability to generalize across diverse problem domains. Thus, this paper proposed a self-adaptive method called GA with Dynamic Strategy Management for the flexibly and efficient construction of maximum LHDs. This method is based on premature convergence prediction, dynamic triggered optimization strategies, and performance control. Furthermore, nearly all critical factor, such as population initialization and selection, crossover, mutation, and local search, are involved in this framework. By comparing this method to LHD construction techniques (Simulated Annealing, Enhanced Stochastic Evolution, and Latin Hypercube Particle Swarm Optimization), as well as the adaptive GAs and state-of-the-art metaheuristics, the algorithm demonstrates superior performance due to its optimized structural self-organization.