Incorporating Decision Maker's Preference Information in Evolutionary Multi-objective Optimization


Abstract

Most optimization problems often involve multiple objectives to be considered simultaneously under some constraints. Unlike single objective problems, the resolution of this kind of problems gives rise to a set of trade-off solutions, called the Pareto front, rather than a single global optimum. During the two last decades, evolutionary algorithms have demonstrated a great success in approximating the whole Pareto front. Recently, researchers have remarked that providing the human decision maker with some hundreds or thousands of optimal solutions makes the decision making task very difficult especially when the number of objectives increases. In reality, since objective functions are not equally important from the decision maker’s viewpoint, this latter is not interested in discovering the whole Pareto front rather than finding only the portion of the front that satisfies his/her preferences which is called the region of interest. For this reason, researchers have mentioned the necessity to hybridize optimization with decision making. The problematic of our PhD thesis is to articulate decision maker’s preferences within multi-objective evolutionary algorithms in order to guide the search towards the region of interest; therefore not only facilitating the decision making task but also saving the computational cost required to explore the remainder of the Pareto frontier. In this research work, we categorise preferences into two main classes. The first class concerns explicit preferences which are expressed in a straightforward manner via one of the available preference modelling tools. The second class concerns implicit preferences which correspond to the desire of exploring special points from the Pareto front in the absence of explicit preferences, i.e., knee regions corresponding to the worthiest regions in terms of trade-offs between the objectives and the nadir point corresponding to the vector composed with the worst objective values at the Pareto optimality stage. Additionally, we consider, in this thesis, the case where there exists more than one decision maker each having his/her own preferences. All the proposed contributions are assessed through experimental studies including comparative experiments against the most prominent recent works by utilizing academic benchmark problems commonly used by the community in addition to a practical instance of the portfolio selection problem.