A Framework for Controllable Multi-objective Learning with Annealed Stein Variational Hypernetworks

Multi-objective optimization (MOO) is a critical area of research in the field of optimization, focusing on problems that involve multiple conflicting objectives. Unlike traditional single-objective optimization, where the goal is to find a single optimal solution, MOO seeks to find a set of solutions that represent the trade-offs between the different objectives. MOO has been successfully applied across diverse fields, from energy system design Marler and Arora (2004) to healthcare treatment planning Craft et al. (2006), demonstrating its versatility in balancing conflicting objectives. In machine learning, multi-objective problems play a significant role in various applications, including recommender systems Milojkovic et al. (2019); Jannach (2022); Zaizi et al. (2023), where they help balance multiple conflicting criteria, and multi-task learning Sener and Koltun (2018); Crawshaw (2020), where they optimize performance across multiple related tasks simultaneously. Furthermore, addressing multi-objective problems in real-world scenarios can be computationally demanding Lin et al. (2022), particularly in high-dimensional settings Nguyen and Tran (2024), as it involves assessing multiple conflicting objectives, which increases computational costs.