Attention-enhanced safe reinforcement learning for metaverse user perception utility optimization in edge metaverse services

Mobile edge computing systems supporting multi-user metaverse services face critical challenges in meeting stringent Motion-to-Photon (MTP) latency requirements below 20 ms while managing head-mounted display energy constraints. Maximizing visual quality while maintaining low latency to enhance user perception utility represents the core optimization challenge. This paper presents a comprehensive user perception utility model combining visual quality perception and MTP latency dimensions, where the visual quality model dynamically adjusts perceptual weights between foreground and background content based on the Weber-Fechner law. We propose an Attention-enhanced Constrained Update Projection Proximal Policy Optimization (ACUP-PPO) algorithm to maximize user perception utility through joint optimization of task offloading, channel assignment, and resource allocation in hierarchical MEC systems with Non-Orthogonal Multiple Access (NOMA) technology. ACUP-PPO employs a three-layer attention mechanism to capture inter-user interference, cross-channel competition, and optimal server selection patterns, while integrating safe reinforcement learning to strictly enforce constraints. Experimental results demonstrate that ACUP-PPO reduces MTP latency by approximately 40% compared to baseline CUP-PPO at 100 MHz bandwidth, achieves 50 fps at 300 MHz (versus 400 MHz for CUP-PPO), and reduces energy consumption by up to 83.7% compared to baseline methods. The algorithm maintains superior user perception utility across diverse metaverse service scenarios, achieving approximately 15% improvement in visual quality-prioritized applications, while exhibiting 99.4% GPU utilization efficiency under high user loads.