Continual Reinforcement Learning-Based Social-Aware Resource Allocation for Uncertain Multi-Modal Virtual-Physical Interaction

In this paper, we study a social-aware resource allocation problem for multi-modal virtual-physical interaction systems. In the considered system, multiple users interact with each other in a shared virtual model (e.g., a Metaverse or digital twin) hosted on an edge server deployed at a base station. This social interaction creates complex dynamics, leading to highly uncertain and time-varying user behaviors and requirements for multi-modal feedback (including visual and haptic). This social-driven uncertainty poses a significant challenge for resource allocation at the edge server. Our objective is to maximize the long-term quality of experience (QoE) of all users by jointly optimizing the edge server's bandwidth and CPU frequency allocations, and multi-modal feedback configurations. To tackle this challenge, leveraging the strong generalization capability of continual reinforcement learning (CRL), we propose a novel CRL-based approach, termed CRL-based social-aware resource allocation approach (CRL-SARA). Simulation results show the effectiveness of CRL-SARA and demonstrate its superiority over the existing schemes.