Bayesian Preference Elicitation for Multiobjective Engineering Design Optimization

Many engineering optimization problems have more than one objective. Although there are methods for identifying the Pareto front of a design space, these methods can be computationally prohibitive, and it is still left to the designer to choose which design on the front to build. Consequently, multiobjective optimization often involves specifying a scalar utility function and then optimizing with respect to this utility function. Preference elicitation algorithms can aid the designer in constructing this utility function. In preference elicitation for engineering design optimization, the ideal algorithm would converge to an accurate utility function with as few preference queries to the designer as possible. This paper extends a prior method for preference realization and tailors it for use in engineering design optimization by deriving a less restrictive inference technique and applying an entropy-based method for posing queries that converges faster to the true global utility function than existing algorithms. This method is then applied to optimization of a next-generation aircraft collision-avoidance system. Finally, an implementation of the algorithm is released as an open-source package in the Julia language.