Inclusion of Preferences in the Design of Evolutionary Optimization Algorithms: An Empirical Study

Abstract

It has already been demonstrated that including the decision maker’s preferences in the design of evolutionary optimization algorithms can help in obtaining non-dominated solutions in the desired regions of the search space or with a desired trade-off between different objectives. In this paper, an empirical investigation on the convergence characteristics of an evolutionary optimization algorithm coupled with a preference model is presented. The paper experimentally shows that inclusion of preferences during the optimization process has the additional benefit of achieving faster convergence on unconstrained multi-objective optimization problems. A multi-objective evolutionary algorithm is modified to include preferences as part of the domination checking procedure. The preference model used in this paper is based on domination cones and is incorporated into the algorithm via a matrix representation that describes the allowed trade-off between objectives. The obtained speed-up is quantified and simulation results are reported on a wide variety of synthetic bi-objective problems gleaned from the literature. The potential advantages of using the preferences during the optimization process are also discussed.