Deep Lifelong Learning for Adaptive Semantic-Aware Content Reuse in UAV-Assisted Metaverse

The vast amount of content generated in the Metaverse and unpredictable user demands make real-time optimization of communication, computing, and caching increasingly challenging. These issues highlight the need for intelligent mechanisms that reduce redundant content transmission and improve resource efficiency. To address this, joint semantic-aware caching and rendering schemes that leverage content similarity are proposed to enable reusability across Metaverse environments. The goal is to optimize user-server associations, caching, and rendering decisions to efficiently utilize network resources, thereby maximizing resource savings and service quality. Reusing content across heterogeneous Metaverse environments, however, requires a learning algorithm capable of adapting to diverse task settings. To this end, a lifelong learning-based algorithm, Deep-Centralized ELLA (DC-ELLA), incorporating dictionary learning is developed to accommodate diverse user requests by dynamically extracting knowledge from different semantic environments. Simulation results show that the proposed caching and rendering schemes significantly outperform traditional approaches, while DC-ELLA enhances convergence speed and stability, demonstrating superior performance in dynamic scenarios. By exploiting knowledge and content from prior requests, the approach achieves scalable adaptation to new Metaverse environments. © 2002-2012 IEEE.