Reinforcement-learning optimal control for type-1 diabetes

摘要

This paper proposes a reinforcement-learning based algorithm for optimal control of blood glucose in patients with type-1 diabetes. Specifically, the algorithm aims to suggest an optimal insulin injection policy. Its performance was assessed using simulations on a combination of the minimum model and part of the Hovorka model. The results show that the proposed methodology successfully regulates and significantly reduces the fluctuation of the blood glucose in both fasting and post-meal scenarios. A comparison between the proposed algorithm and an existing reinforcement learning algorithm also shows the superiority of our method and provide insights on how insulin doses should be chosen.