Deep Q-Learning for Low-Latency Tactile Applications: Microgrid Communications

Abstract

Ultra low-latency is one of the key requirements of tactile internet applications such as microgrid communication, where low end-to-end delay of control messages is essential. In addition, small cell wireless networks emerge as the enabler of networked microgrids, given their coverage, capacity and flexibility. Such densification can help in addressing the stringent QoS requirements of microgrid communications. In dense networks, besides the traditional resource allocation problem, user association can help in reducing communication latency, where users can associate with the base stations that can serve best. In this paper, we jointly address resource allocation and user/device association when Critical User Devices (CUDs), i.e. microgrid controllers, and non-critical users, i.e. User Equipments (UEs), co-exist in a small cell network. We formulate our optimization problem as a resource allocation as well as user association problem. We propose a deep Q-Network based algorithm, namely Delay Minimizing Deep Q-Network (DM-DQN) to address the low-latency requirement. DM-DQN aims at reducing the delay of CUDs by balancing the trade-off between allocating more RBs and associating devices to base stations with high channel quality. We compare the performance of DM-DQN to a tabular Q-learning algorithm. Our results show that the proposed scheme achieves 41% delay reduction for CUDs and it converges faster than Q-learning based scheme algorithm.