Efficient Pruning-Split LSTM Machine Learning Algorithm for Terrestrial-Satellite Edge Network

Abstract

The recent advances in low earth orbit (LEO) satellite-borne edge cloud (SEC) enable resource-limited users to access edge servers via a terrestrial station terminal (TST) for rapid task processing capability. However, the dynamic variation in the TST transmit power challenges the served users to develop optimal computing task processing decisions. In this paper, we propose an efficient pruning-split long short-term memory (LSTM) learning algorithm to address this challenge. First, we present an LSTM algorithm for TST transmit power prediction. The proposed algorithm is then pruned and split to decrease the computing workload and the communication resource consumption considering the limited computing resource of TSTs and served users' quality of service (QoS). Finally, an algorithm split layer selection method is introduced based on the real-time situation of the TST. The simulation results are shown to verify the effectiveness of the proposed pruning-split LSTM algorithm.

Semantic Communication in Satellite-Borne Edge Cloud Network for Computation Offloading

“…where f c is user c's CPU-cycle frequency with the unit cycles/s. The energy required to calculate locally is hence expressed as [1]:…”