New Offline Multi-Objective Optimization Method Boosts Solution Diversity

Multi-objective optimization (MOO) is a powerful tool for problems with several competing goals. However, in offline MOO, where direct function evaluations are impossible, models often struggle with out-of-distribution (OOD) data, leading to inaccurate surrogate models and solutions biased towards the extremes of the Pareto front. To tackle these challenges, a new method called Diversity-driven Offline Multi-Objective Optimization (DOMOO) has been introduced. DOMOO incorporates an accumulative risk control module to estimate and mitigate the OOD risk of candidate solutions, ensuring greater reliability. Furthermore, DOMOO employs a nested Pareto set learning strategy that jointly optimizes preference and Pareto set learning parameters, allowing it to adapt to various Pareto front shapes. It also features a diversity-driven selection strategy using a new indicator, IGD_offline, to extract representative and well-distributed solutions, outperforming existing methods in both convergence and diversity.