Evolutionary Multiobjective Optimization with Humans in the Loop

Roman S?owi?ski (Poznan University of Technology, Poland), Abstract:Evolutionary Multiobjective Optimization (EMO) algorithms have shown remarkable effectiveness in approximating in the objective space the whole set of Pareto-optimal solutions (called Paretofront) to complex multiobjective optimization problems. However, visualizing such a Pareto front is problematic in the case of more than three objectives, hampering the Decision Maker (DM) to choose the solution that ensures the best compromise between the objectives. For this reason, interactive EMO methods have been proposed, involving the DM in the evolutionary search process. In the interactive EMO, preference elicitation phases alternate with optimization phases. In the preference elicitation phase, the DM is asked to express her preferences on a small subset of solutions from the current population. This preference information is converted to DM�s preference model that makes a pressure on the recombination process, directing the evolutionary optimization towards the part of the Pareto front containing the best compromise solution. For interactive EMO with human DM in the loop, it is important the DM understands the impact of her preference information on the direction of optimization convergence. This affects the psychological convergence of the procedure because the DM should become convinced that her preferences built during subsequent iterations are properly represented and used in the optimization phases. We will review several recently proposed interactive EMO methods, focusing on aspects of faithful, efficient, and explainable representations of DM�s preferences.