Modifier-Adaptation Schemes Employing Gaussian Processes and Trust Regions for Real-Time Optimization

Abstract

This paper investigates modifier-adaptation schemes based on Gaussian processes to handle plant-model mismatch in real-time optimization of uncertain processes. Building upon the recent work by Ferreira et al. [European Control Conference, 2018], we present two improved algorithms that rely on trust-region ideas in order to speed-up and robustify the approach. The first variant introduces a conventional trust region on the input variables, whose radius is adjusted based on the Gaussian process predictors' ability to capture the cost and constraint mismatch. The second variant exploits the variance estimates from the Gaussian processes to define multiple trust regions directly on the cost and constraint predictors. These algorithms are demonstrated and compared on a Williams-Otto reactor benchmark problem.