Research Article
A Dynamic Transmission Design via Deep Multi-task Learning for Supporting Multiple Applications in Vehicular Networks
@INPROCEEDINGS{10.1007/978-3-030-99200-2_24, author={Zhixing He and Mengyu Ma and Chao Wang and Fuqiang Liu}, title={A Dynamic Transmission Design via Deep Multi-task Learning for Supporting Multiple Applications in Vehicular Networks}, proceedings={Communications and Networking. 16th EAI International Conference, ChinaCom 2021, Virtual Event, November 21-22, 2021, Proceedings}, proceedings_a={CHINACOM}, year={2022}, month={4}, keywords={Cross-layer transmission design Vehicular communication Multi-task learning}, doi={10.1007/978-3-030-99200-2_24} }- Zhixing He
Mengyu Ma
Chao Wang
Fuqiang Liu
Year: 2022
A Dynamic Transmission Design via Deep Multi-task Learning for Supporting Multiple Applications in Vehicular Networks
CHINACOM
Springer
DOI: 10.1007/978-3-030-99200-2_24
Abstract
We study a cross-layer transmission design problem for vehicular communication networks. Two source-destination links are considered to share the same spectrum resource. Each link intends to send two types of messages to support different delay-sensitive applications. The whole system operates in a dynamic environment in which the small-scale channel fading may change rapidly. Therefore, the sources need to vary their transmission strategies accordingly to efficiently use the available resources while keeping the performance requirements satisfactory. Conventional transmission design via mathematical tools in general demands an iterative computation process and results in high complexity unsuitable for rapid decision-making. In this paper, we propose tackling such a problem by first transforming the transmission design problem into a joint classification-regression problem, and then applying deep multi-task learning (MTL) to solve it. Through simulation results, we show that our method can achieve the similar performance as the transmission design found by mathematical optimizations, with a much faster inference process. The advantages would become even more notable when the network size increases and the environment becomes more complex.
