Chapter 7 - Machine learning–driven sequential optimization using dynamic exploration and exploitation

Abstract

In this chapter, an adaptive surrogate approach for sequential optimization using ensemble machine learning and active learning is presented. This approach, known as active optimizer (ActivO), uses a low variance machine learning model (called the weak learner) to determine promising regions of the design space where the optimum is likely to lie. At the same time, a high-variance machine learning model (called the strong learner) is used to find the exact location of the design optimum, once the promising region has been determined by the weak learner. In addition, criteria for convergence and refinements made to ActivO to enable dynamic exploration-exploitation balance for improved performance are also discussed. Thereafter, ActivO is applied to a canonical optimization problem based on the two-dimensional cosine mixture function consisting of 25 local optima and 1 global optimum. ActivO is compared with five other design optimizers, showing superior rates of convergence and robustness. Subsequently, ActivO is demonstrated for computational fluid dynamics (CFD)-driven optimization of an advanced gasoline compression-ignition engine, with the goal of minimizing the indicated specific fuel consumption while adhering to constraints relating to emissions and operability limits. When compared to a state-of-the-art microgenetic algorithm, ActivO converges to the global optimum significantly faster with 5–7x lower number of compute-intensive CFD simulations, thereby resulting in drastic reduction in computational cost and time-to-design.