Efficient Communication-Computation Tradeoff for Split Computing: A Multi-Tier Deep Reinforcement Learning Approach

Splitting the computation loads of a neural network (NN) training task to multiple stations, split computing has been the most promising technology to sustain high-accuracy model for resource-constrained user equipments (UEs) to empower real-time intelligent services. Nevertheless, different communication link variations and computation capabilities in different stations (including UE and servers) render the overall performance optimization in split computing a critical challenge. In this case, different stations should be able to infer the others' communication/computation capabilities to distributively decide the optimum splitting points of an NN. To this end, in this paper, we propose a multi-tier deep reinforcement learning (DRL) scheme for split computing, by which the UE and edge server can collaboratively and adaptively determine their splitting points and computation resources to optimize the long-term overall training latency through tackling different time-scale sub-optimizations in a sequential manner. With the image recognition task as experimental example, comprehensive simulations are conducted to justify the performances in terms of training latency, model accuracy and energy consumption of the proposed scheme for split computing.