A Generative Adversarial Network Enabled Deep Distributional Reinforcement Learning for Transmission Scheduling in Internet of Vehicles
Faisal Naeem, Sattar Seifollahi, Zhenyu Zhou, Muhammad Tariq
Abstract
The Cognitive Internet of Vehicles (CIoV) is an intelligent network that embeds the cognitive mechanism in the Internet of Vehicles (IoV) to sense the environment and observe the network states to learn the optimal policies adaptively. However, one of the key challenges in CIoV systems is to design a smart agent that can smartly schedule the packet transmission for ultra-reliable low latency communication (URLLC) under extreme random and noisy network conditions. We propose a software defined network (SDN) based scheduling algorithm that leverages generative adversarial network (GAN) based deep distributional Q-network (GAN-DDQN) for learning the action-value distribution for intelligent transmission scheduling. A reward-clipping technique is proposed for stabilizing the training of GAN-DDQN against the effect of broadly spanning utility values. The extensive simulation results verify that GAN-Scheduling achieves higher spectral efficiency (SE), service level agreement (SLA), system throughput, transmission packet rate with lower transmission delay, and power consumption compared to the existing reinforcement learning algorithms.