International Journal of Network Management Компьютерные сети и коммуникации
Accelerating DRL‐Based Task Placement for Immersive Telepresence via Simulation‐Based Training
Аннотация
ABSTRACT Edge‐assisted immersive telepresence can improve application performance by distributing computationally intensive tasks between clients and edge servers. However, effective task placement is challenging because placement decisions must account for heterogeneous computing resources, network conditions, and application‐level quality of experience (QoE). Deep reinforcement learning (DRL) can learn adaptive placement policies, but training directly with actual telepresence workloads requires extensive interactions and makes policy exploration time‐consuming. This paper proposes a profile‐based simulation framework for efficient training of DRL‐based task placement policies for edge‐assisted immersive telepresence. The simulator incorporates measured task execution characteristics, resource contention, and application‐level performance, enabling policy training without repeatedly executing the actual workloads. The trained policy is then evaluated on an edge‐assisted emulation testbed using actual telepresence workloads. Experimental results show that profile‐based simulation enables training of placement policies that improve application performance under stable network conditions and achieve competitive performance under dynamic conditions while maintaining personalization quality. Simulation‐based training also achieves an approximately speedup relative to the estimated minimum execution time in emulation. These results demonstrate the effectiveness of profile‐based simulation for training task placement policies applicable to actual immersive telepresence workloads.