검색 상세

Decentralized Learning Framework for Large-Scale Wireless Sensor Networks

  • 채성윤 (아주대학교 일반대학원)

초록/요약

This thesis targets large-scale sensor networks. First, neighbor discovery is a first step in the initialization of wireless networks in large-scale ad hoc networks. First, we propose a randomized neighbor discovery scheme for wireless networks with a multipacket reception (MPR) capability. We let the nodes to have different advertisement probabilities. The node gradually adjusts its probability according to its operation phases: greedy or slow-start. In the greedy phase, the node advertises aggressively while it does moderately in the slow-start phase. Initial phase and advertisement probability are determined randomly. Then, the nodes change the probability adaptively according to advertisements the reception state from the other nodes. In order to decide the reception state precisely, the exact number of nodes in the network is necessary. To make our proposed scheme work in case of no prior knowledge of the population, we propose a population estimation method based on a maximum likelihood estimation. We evaluate our proposed scheme through numerical analysis and simulation. Through the numerical analysis, we show that the discovery completion time is lower bounded in Θ( Nk ) and upper bounded in Θ( N ln k N ) when there exists N nodes with MPR-k capability. The bounds are the same as those of previous studies that propose static optimal advertisement probability. Through the simulation, we evaluate that our adaptive scheme outperforms in terms of discovery completion time, advertisement efficiency, and wasted time slot ratio than a scheme with static advertisement probability when the population of the network is unknown. Second, The multihop configuration of a large-scale wireless sensor network enables multiple simultaneous transmissions without interference within the network. Existing time division multiple access (TDMA) scheduling schemes exploit gain based on the assumption that the path is optimally determined by a routing protocol. In contrast, our scheme jointly considers routing and scheduling and introduces several new concepts. We model a largescale wireless sensor network as a tiered graph relative to its distance from the sink, and introduce the notion of relay graph and relay factor to direct the next-hop candidates toward the sink fairly and efficiently. The sink develops a transmission and reception schedule for the sensor nodes based on the tiered graph search for a set of nodes that can simultaneously transmit and receive. The resulting schedule eventually allows data from each sensor node to be delivered to the sink. We analyze our scheduling algorithm both numerically and by simulation, and we discuss the impact of protocol parameters. Further, we prove that our scheme is scalable to the number of nodes, from the perspectives of mean channel capacity and maximum number of concurrent transmission nodes. Compared with the existing TDMA scheduling schemes, our scheme shows better performance in network throughput, path length, end-to-end delay, and fairness index. Lastly, we propose an effective neighbor selection scheme for decentralized machine learning algorithm. Existing distributed learning algorithms are defined two approaches, centralized and decentralized. In federated learning, model aggregation process is performed in a centralized server. This centralized architecture has drawbacks, such as communication bottleneck, lack of scalability and sigle-point-of-failure of the server. Gossip learning is representative algorithm in decentralized approach. The performance of gossip learning under nonIID datasets shows low convergence speed and training accuracy. We design a neighbor selection scheme that each collector node enables to select the best neighbor set for model aggregation. To cope this challenge, we use statistical distance measure. Then, we propose more efficient algorithm by reducing neighbor selection frequency. We develop a simulator and measure performances in terms of training accuracy level, convergence speed. Simulation results show the proposed algorithm significantly performs compare to federated learning with full and partial participation and gossip learning.

more

목차

Chapter I Introduction 1
Chapter II System Architecture 4
II.A Neighbor discovery 4
II.A Transmission scheduling 6
II.A Decentralized machine learning 7
Chapter III Neighbor discovery 8
III.A Neighbor discovery protocols 11
III.B Proposed Scheme 14
III.C Performance Analysis 23
III.D Simulation Result 28
III.E Chapter summary 36
Chapter IV Transmission scheduling 38
IV.A Related work 41
IV.B Proposed scheme 44
IV.C Performance Evaluation 58
IV.D Chapter summary 75
Chapter V Decentralized learning algorithm 78
V.A Introduction 78
V.B Related Works 80
V.C System Model 82
V.D Decentralized Learning Algorithm 86
V.E Performance Evaluation 95
V.F Chapter Summary 105
Chapter VI Conclusions 106
Bibliography 109

more
반출 Meta View 목록