Computationally Efficient Bayesian Optimization for Multi-objective Industrial Applications

Piyush Pandita,Sayan Ghosh,Liping Wang

DOI: https://doi.org/10.2514/6.2021-1482

2021-01-01

Abstract:Industrial problems often take the form of expensive laboratory experiments or computationally-heavy simulations. These expensive experiments or simulations provide valuable information about the underlying latent physical process being modeled. One is often faced with the problem of optimizing the quantity of interest (QoI) being represented by the information source. Querying the information source tens of thousands of times, generally required by state-of-the-art optimization algorithms, is practically infeasible. The problem becomes cumbersome when the number of objectives being optimized is greater than one. In such multi-objective optimization (MOO) scenarios, expected improvement (EI)-based methods have demonstrated great efficiency in converging to the optimal solution while taking into consideration a fixed budget. The EI-based methods, used for expensive MOO problems, are broadly called EI-hypervolume (EIHV) methods. The EI-based algorithm called, Bayesian global optimization (BGO), builds cheap probabilistic models of the objectives, using Gaussian process regression(GPR), and selects the next input to query the information source from amongst a finite set of candidate inputs. The EIHV weighs the potential improvement associated with each of the candidate designs. This iterative process continues until one exhausts the budget. The evaluation of the EIHV becomes painfully expensive, scales exponentially, as the number of objectives increases. The finite set of possible inputs, used to evaluate the EIHV and select the next query, ceases to be a computationally feasible option in order for the overall optimization to proceed. In this work, we aim to investigate the performance of an EI-based algorithm, for optimizing the EIHV for MOO problems. At each iteration of the optimization problem, we optimize the EIHV using EI-based BGO, in order to be computationally efficient. We demonstrate the approach on two mathematical MOO problems and one challenging industrial problem with four objectives. We also compare the results in terms of convergence and savings in computational time with the finite-grid based approach.