A Deep Q-Learning Approach for an Efficient Resource Management in Vehicle-to-Everything Slicing Environment

Abstract

In the context of 5G vehicle-to-everything (V2X) communications, network slicing has been presented as a prominent solution to enable the coexistence of different V2X use cases within the same infrastructure. Among the main challenges in V2X network slicing is proposing an appropriate resource management approach that allows resources to be used efficiently while protecting the isolation between slices. One of the benchmarking resource management approaches is the static allocation that provides a full isolation between the slices. This static allocation has serious limitations when dealing with dynamic scenarios, particularly when some slices are overloaded. This paper proposes a deep Q-learning approach to readjust static resource allocation by enabling an opportunistic sharing mechanism in case there is an overloaded slice in the system. More specifically, the main idea of our proposal is to extract an appropriate amount of resources from each available slice within the same infrastructure when receiving rejected connection requests from overloaded slices, while taking into account the isolation aspect of the slicing environment. Simulation results showed better performance in terms of maximizing the use of resources, reducing the probability of new calls being blocked and the probability of handover dropping in the system, while maintaining high isolation compared with other solutions.