Energy-Efficient Topology Construction via Power Allocation for Decentralized Learning via Smart Devices With Edge Computing

Data privacy preservation has drawn much attention in emerging machine learning applications. Decentralized learning among smart devices over wireless networks is thus developed to guarantee data security and eliminate the involvement of parameter servers to avoid transmission bottlenecks. However, the previous research focuses on data compression and exchange rules of model parameters among smart devices. Still, it neglects the interplay between link cardinality, transmission power consumption, and collision in transmission. To jointly optimize these issues, in this paper, we first set collision and interference aside and formulate a new optimization problem, named GreenDL, and then extend GreenDL to be collision-aware, namely, GreenDL-CA, by restricting the maximum degree of each smart devices. We prove their hardness and propose two approximation algorithms dubbed as CoTRAIN and CoTRAIN-CA for GreenDL and GreenDL-CA, respectively. Experiment and simulation results manifest that both CoTRAIN and CoTRAIN-CA reduce more than 20% power compared with the other heuristics without sacrificing the convergence rate in the decentralized learning practices.