In the present study, we investigated the effect of different reward functions in insulin regulation using reinforcement learning. An artificial pancreas system is able to deliver insulin into the body in an automated way. The control algorithm of an automated insulin delivery system is a key player in achieving personalized therapy. Neural networks provide an approach to customize insulin administration by learning the patient’s habits and administering insulin accordingly. Therefore, we conducted experiments with neural networks based on reinforcement learning. Our goal was to find a neural network-based model and reward function that could learn the patient’s behavior and administers insulin with the best time in ranges. We evaluated the method on simulated virtual patients when sensor noise occurs. The results show that the bump functions were the most efficient in providing acceptable time in ranges.