Improving Pareto Front Learning via Multi-Head HyperNetwork

Multi-objective optimization (MOO) plays a critical role in many machine learning applications. The set of optimal solutions for these MOO problems are referred to as the Pareto front, in which each point on the front represents different trade-off between conflicting objectives. Pareto front learning is recently introduced as an effective approach to allow one to flexibly derive the optimal solution given the desired trade-off after training. This setup is practical as in many scenarios, the decision-makers can not specify the preference of one Pareto solution over another, and must switch between them depending on the situation.

Although Pareto front Learning offers many advantages in approximating the true Pareto front, existing methods are achieving the solutions that are either suboptimal or not widely dispersed. To overcome this issue, we propose a novel Multi-head HyperNetwork (MHN) architecture to improve the quality of the obtained Pareto front. Particularly, we generate multiple Pareto solutions from a set of diverse trade-off preferences via the MHN and improve the quality of the Pareto front by maximizing the hypervolume value defined by these solutions. The experimental results in a wide variety of machine learning tasks show that the proposed method significantly outperforms the baselines in producing high-quality Pareto front.