Deep reinforcement learning-based resource reservation algorithm for emergency Internet-of-things slice

Based on the requirements of ultra-low latency services for emergency Internet-of-things (EIoT) applications, a multi-slice network architecture for ultra-low latency emergency IoT was designed, and a general methodology framework based on resource reservation, sharing and isolation for multiple slices was proposed. In the proposed framework, real-time and automatic inter-slice resource demand prediction and allocation were realized based on deep reinforcement learning (DRL), while intra-slice user resource allocation was modeled as a shape-based 2-dimension packing problem and solved with a heuristic numerical algorithm, so that intra-slice resource customization was achieved. Simulation results show that the resource reservation-based method enable EIoT slices to explicitly reserve resources, provide a better security isolation level, and DRL could guarantee accuracy and real-time updates of resource reservations. Compared with four existing algorithms, dueling deep Q-network (DQN) performes better than the benchmarks. © 2020, Editorial Board of Journal on Communications. All right reserved.