A New Linear Programming Approach and a New Backtracking Strategy for Multiple-Gradient Descent in Multi-Objective Optimization

TL;DR

This paper introduces a novel linear programming-based method for shared descent directions in multi-objective optimization and a new backtracking strategy to enhance Pareto set exploration.

Contribution

It presents a new LP approach for finding shared descent directions and a backtracking strategy to improve Pareto front exploration in multi-objective optimization.

Findings

The LP-based method effectively finds shared descent directions.

The backtracking strategy enhances Pareto front exploration.

Theoretical analysis confirms the method's properties.

Abstract

In this work, the author presents a novel method for finding descent directions shared by two or more differentiable functions defined on the same unconstrained domain space. Then, the author illustrates an alternative Multiple-Gradient Descent procedure for Multi-Objective Optimization problems that is based on this new method. In particular, the proposed method consists in finding the shared descent direction solving a relatively cheap Linear Programming (LP) problem, where the LP's objective function and the constraints are defined by the gradients of the objective functions of the Multi-Objective Optimization problem. More precisely, the formulation of the LP problem is such that, if a shared descent direction does not exist for the objective functions, but a non-ascent direction for all the objectives does, the LP problem returns the latter. Moreover, the author defines a new backtracking strategy for Multiple-Gradient Descent methods such that, if the proposed LP is used for computing the direction, the ability to reach and/or explore the Pareto set and the Pareto front is improved. A theoretical analysis of the properties of the new methods is performed, and tests on classic Multi-Objective Optimization problems are proposed to assess their goodness.