Original title: Multi-Objective Optimization via Wasserstein-Fisher-Rao Gradient Flow
Authors: Yinuo Ren, Tesi Xiao, Tanmay Gangwani, Anshuka Rangi, Holakou Rahmanian, Lexing Ying, Subhajit Sanyal
The article delves into optimizing multiple conflicting goals, a task known as Multi-objective optimization (MOO). Inspired by molecular dynamics, the team devises a new method using particle interactions. This technique, blending Langevin and birth-death dynamics, includes a “dominance potential” guiding particles toward global Pareto optimality. Unlike previous methods, this one can relocate dominated particles, making it adept at handling complex Pareto fronts. The approach is grounded in theory as a Wasserstein-Fisher-Rao gradient flow, ensuring convergence. Tests prove its superiority over existing methods across tough synthetic and real-world datasets. Essentially, it’s a smarter way to handle conflicting goals, like juggling multiple objectives, ensuring they’re all optimized effectively even when they compete with each other.
Original article: https://arxiv.org/abs/2311.13159