Deep Learning-Based Resource Allocation Scheme For Vehicle to Everything Communication

Abstract

Vehicle to Everything (V2X) is rapidly evolving with the development of 5G technology. V2X will provide prominent services, including vehicle to vehicle communication (V2V) and vehicle to network infrastructure (V2I). It would become possible that not only urgent signal transmission to announce an emergency, but also multimedia data transfer as a streaming video service or vision data for autonomous driving. That services usually require a high data rate, reliability , and low-latency. To issue the requirements, combining various techniques is considered, such as Non-orthogonal multiple access (NOMA), Heterogeneous network architecture (Hetnet), and full-duplex. However, complexity would be a significant obstacle to utilize the combined techniques in a system. Specifically, we target on two challenges from the complexity of the combined system: reducing the control burden by using a distributed method and allocating resources on the environment with a high action space.

The control burden means that the central node has to control many devices. It includes a problem to use crowd control channels. To allocate resources, channel information usually has to be exchanged in V2X. A central node must simultaneously perform a series of processes involving complex calculations and the exchange of various information on numerous devices. To reduce the control burden, we propose an autonomous power allocation scheme based on deep learning to reduce the control burden. By using our distributed deep learning architecture, devices can determine to transmit power without any involvement of the other devices. For the device to device communication (D2D), it can achieve near-optimal cell throughput while maintaining constraints of interference to base stations.

The high action space problem arises when different kinds of resources affect each other and need to be optimized for the same objective. Because of the combination of resources, there are countless cases of actions. It is becoming almost impossible to fully explore these action spaces to optimize overall performance. To issue the problem, we propose a deep genetic algorithm for a NOMA-V2X resource allocation problem. Deep learning reduces dimensions continuously while allocating resources using a genetic algorithm. As a result, it can achieve 20\% better throughput than using a genetic algorithm only.

Based on deep learning, we overcome the technical limitations: realizing fully distributed and high action space resource allocation. The proposed schemes can be applied in the future network system, which is becoming highly complex.