Probabilistic optimal control of blood glucose under uncertainty

Type 1 diabetes mellitus is a chronic disease requiring careful management of insulin infusion rates. A novel simulation-based approach to probabilistic optimal control of blood glucose concentration using Gaussian Process Dynamic Programming (GPDP) and Bayesian active learning is proposed. GPDP is an approximate value function method that integrates reinforcement learning with Gaussian Processes (GP) for seeking an optimal control policy in the face of an uncertain dynamics. The obtained control policy is compactly represented the hyper-parameters (mean and variance) of a Gaussian process over a wide range of physiological states to facilitate an "on-a-chip" implementation. Safe adaptation of the simulation-based control policy to a patient-specific lifestyle upon data is presented. Simulation results are very promising.