Project R-14165

Bayesian many-objective optimization with noise and user preference learning. (Research)

Optimization problems in practice are often analyzed using computer models (e.g., stochastic simulation models). These models can be expensive, though, in terms of computation time and/or cost. To remedy this issue, metamodels can be used, which yield (data-driven) approximations of the simulation model, while being computationally cheap(er) to evaluate. Machine Learning (ML)-based optimization is a relatively novel, yet highly promising field within operations research. This project particularly focuses on ML-based many-objective optimization of (noisy) conflicting objectives, using preference learning. In many-objective settings, the decision-maker (DM) is typically only interested in a subset of the Pareto optimal solutions; hence, incorporating these user preferences in the optimization framework would not only allow the algorithm to converge to the relevant solutions faster, but also to scale up to larger sets of objectives in a data-efficient manner. Learning these underlying user preferences is challenging, though, as preference information is typically distorted by noise (e.g., due to cognitive burden of the DM). The key research objectives of this project are threefold: (i) to study the scalability issue in many-objective optimization; (ii) to develop a methodology for modelling noisy preferences, and to use this information in an ML-based optimization framework ; and (iii) to develop novel performance metrics for measuring the quality of the obtained solutions.

01 October 2023 - 30 September 2026