Deep Reinforcement Learning for Closed-Loop Blood Glucose Control: Two Approaches

Reinforcement learning, thanks to the observation-action approach, represents a useful control tool, in particular when the dynamics are characterized by strong non-linearity and complexity. In this sense, it has a natural application in the biological systems field where the complexity of the dynamics makes the automatic control particularly challenging. This paper presents a combined application of neural networks and reinforcement learning, in the so-called field of deep reinforcement learning, for the glucose regulation problem in patients with diabetes mellitus. The glucose control problem is solved through the Deep Deterministic Policy Gradient (DDPG) and the Soft Actor-Critic (SAC) algorithms, where the environment exploited for the agent's interactions is represented by a glucose model that is completely unknown to agents. Preliminary results show that the DDPG and SAC agents can suitably control the glucose dynamics, making the proposed approach promising for further investigations. The comparison between the two agents shows a better behaviour of SAC algorithm.