DRL-based latency-energy offloading optimization strategy in wireless VR networks with edge computing

Abstract

The increase in data paths and the resulting latency growth in Wireless Virtual Reality (WVR) can significantly affect user experience. Mobile Edge Computing emerges as an effective solution to address these issues. However, offloading methods based on Deep Reinforcement Learning (DRL) face hurdles like limited environmental exploration and prolonged user waiting time. To address the mentioned challenges in WVR edge computing, where computational offloading involves multiple devices and edge servers, we aim to minimize system latency and reduce energy consumption. Therefore, we introduce the Task Prediction and Multi-objective Optimization Algorithm (TPMOA). First, we reduce the time users wait for rendering results by predicting their viewpoints. Next, we apply an entropy-innovated DRL algorithm to the latent space for computation offloading. Through representation learning, we establish a reward function that includes latent objectives and optimizes the experience replay buffer. This approach allows us to train and select the optimal offloading strategy, thereby reducing rendering latency and system energy consumption. Our experiments show that our approach effectively tackles the challenges of limited environmental exploration ability and extended user waiting time. Specifically, our method outperforms the RNN-based AC method significantly, reducing latency by 11.39%.