Fault Tolerant Control of Blood Glucose Concentration Using Reinforcement Learning

In this paper, the main focus is on blood glucose level control and the possible sensor and actuator faults which can be observed in a given system. To this aim, the eligibility traces algorithm (a Reinforcement Learning method) and its combination with sliding mode controllers is used to determine the injection dosage. Through this method, the optimal dosage will be determined to be injected to the patient in order to decrease the side effects of the drug. To detect the fault in the system, residual calculation techniques are utilized. To calculate the residual, it is required to predict states of the normal system at each time step, for which, the Radial Basis Function neural network is used. The proposed method is compared with another reinforcement learning method (Actor-Critic method) with its combination with the sliding mode controller. Finally, both RL-based methods are compared with a combinatory method, neural network and sliding mode control. Simulation results have revealed that the eligibility traces algorithm and actor-critic method can control the blood glucose concentration and the desired value can be reached, in the presence of the fault. However, in addition to the reduced injected dosage, the eligibility traces algorithm can provide lower variations about the desired value. The reduced injected dosage will result in the mitigated side effects, which will have considerable advantages for diabetic patients.