Focusing Search in Multiobjective Evolutionary Optimization through Preference Learning from User Feedback

The problem of computing the set of Pareto-optimal solutions in multiobjective optimization has been tackled by means of different approaches in previous years, including evolutionary algorithms. A key advantage of computing the whole set of Pareto-optimal solutions is completeness: None of the solutions that might be maximally preferred by the user is lost. An obvious disadvantage, however, is the computational effort, which may become prohibitive for high-dimensional problems. Besides, the resulting set itself may become rather large, making it impracticable to present the entire set to the user. In order to mitigate these problems, we propose a method for incorporating user-feedback in the optimization process by asking the user for pairwise preferences. The pairwise preferences thus produced are then used as training examples for a preference learning method leading to a hypothetical utility model that approximate the true but unknown (latent) utility function of the user. This model is used to focus the search on the most promising parts of the Pareto front, which is approximated by an evolutionary algorithm.