Non-linear Continuous Action Spaces for Reinforcement Learning in Type 1 Diabetes

Abstract

AbstractArtificial Pancreas Systems (APS) aim to improve glucose regulation and relieve people with Type 1 Diabetes (T1D) from the cognitive burden of ongoing disease management. They combine continuous glucose monitoring and control algorithms for automatic insulin administration to maintain glucose homeostasis. The estimation of an appropriate control action—or—insulin infusion rate is a complex optimisation problem for which Reinforcement Learning (RL) algorithms are currently being explored due to their performance capabilities in complex, uncertain environments. However, insulin requirements vary markedly according to sleep patterns, meal and exercise events. Hence, a large dynamic range of insulin infusion rates is required necessitating a large continuous action space which is challenging for RL algorithms. In this study, we introduced the use of non-linear continuous action spaces as a method to tackle the problem of efficiently exploring the large dynamic range of insulin towards learning effective control policies. Three non-linear action space formulations inspired by clinical patterns of insulin delivery were explored and analysed based on their impact to performance and efficiency in learning. We implemented a state-of-the-art RL algorithm and evaluated the performance of the proposed action spaces in-silico using an open-source T1D simulator based on the UVA/Padova 2008 model. The proposed exponential action space achieved a 24% performance improvement over the linear action space commonly used in practice, while portraying fast and steady learning. The proposed action space formulation has the potential to enhance the performance of RL algorithms for APS.KeywordsReinforcement learningGlucose regulationContinuous action space