The paper discusses a problem called MO optimization, which involves finding solutions that balance multiple, often conflicting objectives. For example, in e-commerce, a recommendation system may need to balance between maximizing revenue and customer satisfaction. The paper presents a new method for training neural networks that enables a posteriori MO decision-making and demonstrates its effectiveness in various real-world problems.
Novel approach
The authors propose a novel approach to training neural networks for multi-objective decision-making, where the neural networks generate multiple solutions that span and uniformly cover the Pareto front, which is the set of optimal trade-off solutions. Unlike traditional approaches that require the user to specify the trade-off vectors beforehand, the proposed approach does not require this information and enables a posteriori MO decision-making. The approach uses maximization of hypervolume (HV), which is a metric that measures the quality and diversity of a set of solutions, to train the neural networks.
Well-spread outputs
Experiments conducted on different multi-objective problems show that the proposed approach returns well-spread outputs across different trade-offs on the approximated Pareto front. Especially in cases that are challenging and complex because of an asymmetric Pareto front (when one objective's improvement comes at a high cost to the improvement of other objectives).
EMO 2023 is the 12th Edition of the biannual International Conference Series on Evolutionary Multi-Criterion Optimization (EMO). It focuses on solving real-world problems in government, business, and industry.
More information:
- The paper is called 'Multi-objective Learning Using HV Maximization'. Authors: Timo M. Deist (Former CWI employee), Monika Grewal (CWI), Peter A. N. Bosman (CWI/Delft University of Technology), Frank J. W. M. Dankers and Tanja Alderliesten (Leiden University Medical Center). It is published in the proceedings of the EMO 2023 conference.
- A preprint is available at https://arxiv.org/abs/2102.04523.
- The source code is available at https://github.com/timodeist/multi_objective_learning.
