Partitioned Edge Learning Over Fast Fading Channels

Abstract

The implementation of the partitioned edge learning (PARTEL) framework in practical fast-fading wireless systems with time-varying channels is investigated in this paper. By exploiting the benefit of only training and transmitting a light-size sub-model on devices, PARTEL enjoys both enhanced computation and communication efficiency compared to federated edge learning framework where each device needs to train and transmit all model parameters. In this work, we aim at further enhancing the learning efficiency via minimizing the training latency of each training round in practical fast fading channels, where each round includes multiple channel time-coherence time durations. The challenges arise from the unknown channel state information (CSI) of future durations and the coupling between load balancing and bandwidth allocation among devices. To this end, an equivalent Markov decision process (MDP) problem is derived, where each decision step corresponds to one channel coherence-time duration. The learning load balancing is first determined based on an existing design for static channels by using the expected channel gains. Then, the bandwidth allocation for uploading the local sub-model updates in each duration is sequentially determined by the proposed optimization algorithm. The spectrum efficiency is enhanced since the bandwidth is adaptively allocated to all devices according to their communication and computation statuses in each coherence-time duration. Finally, extensive simulations are conducted to show the superiority of our proposed algorithms over the existing benchmarks. Specifically, the proposed algorithm can reduce the one-round latency by up to 25.51% for the bandwidth of 70 MHz.