Deep Reinforcement Learning (DRL) for resource allocation in Cloud: Review and Prospects

Deep Reinforcement Learning (DRL) is robust to trend changes in dynamic cloud environments. The combination of Deep Neural Networks for learning the Q values in Reinforcement Learning with regular updates to the trained model shows promising results in reducing makespan and energy consumption. Vanilla DRL and various schemes that address shortcomings in the basic DRL are analyzed for the scheduling of tasks in cloud. The design parameters of the neural network and reinforcement learning are compared to present configurations that work well for a particular cloud environment. The advantages and shortcomings of the approaches are highlighted and scrutinized compared to several baseline algorithms. The OpenAI Gym package is discussed to give novice researchers insight into how to start work on DRLs. The prospects of applying DRL in cloud scheduling are put forth for researchers to construct efficient learning networks for efficient task scheduling.