Improving Mobile Edge Computing Orchestration Using Deep Reinforcement Learning Through Environment Predictions

Edge Computing (EC) in Hospital and Healthcare environments require an efficient offloading orchestration between Internet of Things (IoT) sensors, mobile smart devices, and the network edge, since many of their applications (APP) have low latency constraints that are linked to critical health and safety demands. This work focuses on using environment predictions in a a novel offloading orchestration algorithm based on Deep Reinforcement Learning (DRL) that we have introduced before. The predictions are obtained by combining Long Short Memory (LSTM) with three Deep Neural Networks (DNN), to assess APP's tasks dependency, predict possible next tasks, estimate future CPU availability of edge nodes and predict the mobility path of devices. We test our solution in a dynamic EC Hospital environment with different IoT and mobile devices. Tests show that the addition of LSTM with several DNNs significantly improves the DRL orchestration algorithm, in terms of task completion within the delay requirements and in Quality of Experience (QoE), compared to other offloading solutions from the literature.