CAREER: Data-Enabled Neural Multi-Step Predictive Control (DeMuSPc): a Learning-Based Predictive and Adaptive Control Approach for Complex Nonlinear Systems

University Of Houston, Houston TX

Investigators

Abstract

This Faculty Early Career Development (CAREER) grant will fund research that enables new knowledge related to a data-enabled automatic control approach for complex processes which are hard to describe from first principles and are changing over time, promoting the progress of science and advancing national health and prosperity. Among all the processes with the above outlined properties, one compelling example is represented by the regulation of blood glucose in people with type 1 diabetes by means of exogenous insulin. Insulin therapy is affected by a considerable number of unknown and hidden physiological variables that change with patient’s lifestyle and growth/aging, requiring frequent interventions by patients and their caregivers. Despite intensive and burdensome treatment, the majority of patients still fail to meet their prescribed glycemic targets leading to complications which are costly to both the individual and the healthcare system. This project will support fundamental research to provide needed knowledge for the development of data-driven and learning-based predictive and adaptive automatic control. The success of this project will enable a framework for optimal regulation and adaptation to changes with application in healthcare, biomedical, advanced manufacturing, chemical or automotive industries. The research is integrated with educational and outreach activities to broaden participation of groups traditionally underrepresented in control research and contribute positively to engineering education. This research aims to make fundamental contributions to data-enabled predictive and adaptive control to overcome several limitations affecting existing predictive control approaches, including large errors in the model predictions for long prediction horizons due to large plant-model mismatch and unmodeled dynamics, as well as policy parameters that are static and do not adapt to varying operating conditions. The project will (1) exploit the use of multi-step ahead output predictors with a structure nonlinear in the state and affine in the future control moves, (2) identify the unknown mappings in the predictor parameterizations from input-output data by means of neural networks embedding prescribed behavioral guarantees in their structure, (3) integrate the predictors into a linear time-varying model predictive control framework, and (4) use Bayesian Optimization to tune and adapt the parameters of the controller to changes in the dynamics. The algorithms will be validated on the motivating examples of automated glucose regulation in people with type 1 diabetes by performing extensive in-silico trials with a metabolic simulator. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.