An efficient near-optimal Latin hypercube design for large datasets

As a common method for design and analysis of computer experiments (DACE), optimal Latin hypercube designs (OLHD) offer excellent space-filling properties but are computationally intensive with larger datasets. Techniques like translational propagational LHD (TPLHD) and improved translational propagational algorithm (iTPA) that compose near-optimal LHDs reduce computational cost with adequate space-filling properties but struggle with large, high-dimensional datasets as well. This study introduces cyclic latin hypercube design (CLHD) to address these challenges by arranging elements at regular intervals, applying a roll operation, and utilizing an enhanced stochastic evolutionary (ESE) algorithm for efficient near-optimal LHD construction. Performance comparisons with OLHD, iTPA, random LHD, and periodic LHD across various dataset sizes show that CLHD successfully reduces computational time and enhances space-filling properties, up to 49.4 % better than random LHD, 30.0 % better than iTPA, and 4.2 % better than periodic LHD. Consequently, CLHD is an efficient alternative for large, high-dimensional datasets where OLHD is infeasible.