Evolution Assisted Pareto Set Learning for Multi-Modal Multi-Objective Optimization

We propose a novel framework that integrates multi-modal multi-objective optimization evolutionary algorithms (MMOEAs) with Pareto Set Learning (PSL) to more effectively address multi-modal multi-objective optimization problems (MMOPs), which are characterized by multiple Pareto sets corresponding to the same Pareto front. In the proposed framework, evolutionary search is first employed to obtain a diverse set of initial solutions. These solutions are then used to construct a dataset of preference-solution-label triplets, which guides the pre-training of a Pareto set model. Following this stage, the model is further fine-tuned using the PSL methodology. To better capture the diversity of solutions in the decision space, we adopt a backbone network with multiple output heads. The resulting model is capable of generating multiple continuous Pareto sets. Preliminary experiments on benchmark MMOPs show that our method can successfully learn all major Pareto sets, outperforming both PSL and its improved variant in terms of decision-space diversity and coverage completeness. Code is available at: https://github.com/fortherace/EAPSL.