Deep Neural Network Architectures for an Embedded MPC Implementation: Application to an Automated Insulin Delivery System

Abstract

Computing a Model Predictive Controller (MPC) requires high computational loads, which typically challenges its implementation in embedded hardware. Recently, learning MPCs through Neural Networks (NNs) has been suggested as suitable methodology for on-chip MPC implementations. In this paper, we assess the performance of this methodology by training two different NN architectures for learning the University of Virginia's RocketAP system, a clinically validated Automated Insulin Delivery (AID) algorithm that contains at its core an individualizable MPC with adaptive weights. The work has two main motivations. The first one is to find a suitable path for an embedded implementation of the above-mentioned AID system in insulin pumps or wearables. The second one is to report the results of how this methodology of learning MPCs with NNs performs in a clinically validated MPC that is significantly more complex than the previously reported use cases. The results indicate strong capabilities of NNs for efficient learning of this MPC, achieving a 99.7% of accuracy while requiring a small memory footprint on the order of kilobytes (kB). We also show that deep residual neural network architectures may be a better choice for this type of scenario.