Optimal Linear Control of Blood Glucose

One criticism of neural network controllers (neuro-controllers) is that the analytical model of the controller is not defined; therefore contemporary optimization techniques in control systems cannot be applied to the closed loop system. Often control parameters are tuned online because of inaccuracies due to linearity assumptions and reduction of order. This paper demonstrates how the specialized learning technique can be applied to develop an optimal controller which does not require additional online tuning even when the process model is a complex one such as the blood glucose control system for a Type I diabetic patient. The system has been modeled using the linear quadratic regulator (LQR) technique to ensure optimal control and then used to train the neuro-controller via the specialized learning technique. The result is an optimal neuro-controller which controls the blood glucose system in a Type I diabetic patient, even in the presence of large disturbances.