Abstract
Reinforcement learning has been increasingly applied in monitoringapplications because of its ability to learn from previous experiences and canmake adaptive decisions. However, existing machine learning-based healthmonitoring applications are mostly supervised learning algorithms, trained onlabels and they cannot make adaptive decisions in an uncertain complexenvironment. This study proposes a novel and generic system, predictive deepreinforcement learning (PDRL) with multiple RL agents in a time seriesforecasting environment. The proposed generic framework accommodates virtualDeep Q Network (DQN) agents to monitor predicted future states of a complexenvironment with a well-defined reward policy so that the agent learns existingknowledge while maximizing their rewards. In the evaluation process of theproposed framework, three DRL agents were deployed to monitor a subject'sfuture heart rate, respiration, and temperature predicted using a BiLSTM model.With each iteration, the three agents were able to learn the associatedpatterns and their cumulative rewards gradually increased. It outperformed thebaseline models for all three monitoring agents. The proposed PDRL framework isable to achieve state-of-the-art performance in the time series forecastingprocess. The proposed DRL agents and deep learning model in the PDRL frameworkare customized to implement the transfer learning in other forecastingapplications like traffic and weather and monitor their states. The PDRLframework is able to learn the future states of the traffic and weatherforecasting and the cumulative rewards are gradually increasing over eachepisode.