Decision Variable Learning
Matheus C. dos Santos, Joel Alves de Oliveira, André Britto · 2019
Many alternatives to traditional Multi-Objective Optimization Algorithms are emerging due to the increasing number of Multi-Objective Problems with a high degree of complexity, such as Many-Objective Problems. Among these alternatives, the methods known as surrogate stand out. These methods seek to construct new models for the objective functions based on the data obtained previously from the actual objective functions. The input of these models are a set of vectors in decision variable space and the output are the values of the objective functions. In this work, the Decision Variable Learning (DVL) algorithm is proposed, which presents an inverse idea to traditional surrogates. In the DVL, machine learning models will be used to learn the behavior of the decision variables in Many-Objective Optimization Problems. In this context, we have enough information to learn from the objective vectors, to predict near optimal solution in decision variable space. DVL algorithm will be evaluated using benchmark problems and its results will be compared with the NSGA-III algorithm.