Dr. Jesse Sestito
In engineering design, it is commonplace to modify design parameters such that a set of properties or attributes are optimized for a particular application. This optimization process can be successfully performed using optimization techniques built on machine learning such as sequential multi-objective Bayesian optimization (S-MOBO). S-MOBO takes in existing sets of design parameters and their corresponding solutions and recommends design parameters that are most likely to produce a solution on the Pareto front, the set of non-dominated solutions. Though S-MOBO is a powerful technique, only one iteration of the design can be built at a time. In design problems that can be parallelized, batch multi-objective Bayesian optimization (B-MOBO) can be used instead to accelerate the optimization process by recommending multiple sets of design parameters whose solutions are expected to exist along the Pareto front. Though this process increases overall computation time, the real-time computation is reduced. In this work, we develop a new B-MOBO method as well as a Python framework to support any MOBO method. The new B-MOBO method is developed using a gaussian process model surrogate to inform the new acquisition function. The new B-MOBO method will more efficiently recommend batch samples which target the Pareto front, reducing overall real-time computation.
Holder, Adelle and DeBruin, Henry, "Development of an Efficient Batch Multi-objective Bayesian Optimization Method for Engineering Design" (2022). Symposium on Undergraduate Research and Creative Expression (SOURCE). 1083.