Learning-Based Multi-Tier Split Computing for Efficient Convergence of Communication and Computation

With promising benefits of splitting deep neural network (DNN) computation loads to the edge server, split computing has been a novel paradigm achieving high-quality artificial intelligence (AI) services for energy-constrained user equipments (UEs). To satisfy the service demands of a large number of UEs, traditional edge-UE split computing evolves toward multi-tier split computing involving edge and cloud servers with different capabilities, leading to a “complex” optimization involving communication and computation. To tackle this challenge, in this paper, we propose a multi-tier deep reinforcement learning (DRL) decision-making scheme for distributed splitting point selection and computing resource allocation in three-tier UE-edge-cloud split computing systems. With the proposed scheme, the high-dimensional optimization can be tackled by UEs and an edge server with different control cycles through performing local decision-making tasks in a sequential manner. Based on the policies updated by UEs and edge server in successive stages, the overall performance of split computing can be continuously improved, which is justified through a theoretical convergence performance analysis. Comprehensive simulation studies show that the proposed multi-tier DRL decision-making scheme outperforms conventional split computing schemes in terms of the overall latency, inference accuracy and energy efficiency, to practice multi-tier split computing.