摘要:
Distributed optimization is a powerful paradigm to solve various problems in machine learning over networked systems. Existing first-order optimization methods perform cheap gradient descent by exchanging information per iteration only with single-hop neighbours in a network. However, in many agent networks such as sensor and robotic networks, it is prevalent that each agent can interact with other agents over multi-hop communication. Therefore, distributed optimization method over multi-hop networks is an important but overlooked topic that clearly needs to be developed. Motivated by this observation, in this paper, we apply multi-hop transmission to the outstanding distributed gradient descent (DGD) method and propose two typical versions (i.e., consensus and diffusion) of multi-hop DGD method, which we call CM-DGD and DM-DGD, respectively. Theoretically, we present the convergence guarantee of the proposed methods under mild assumptions. Moreover, we confirm that multi-hop strategy results in higher probability to improve the spectral gap of the underlying network, which has been shown to be a critical factor improving performance of distributed optimizations, thus achieves better convergence metrics. Experimentally, two distributed machine learning problems are picked to verify the theoretical analysis and show the effectiveness of CM-DGD and DM-DGD by using synthesized and real data sets.
期刊:
IEEE Open Journal of the Computer Society,2021年2:393-406
作者机构:
[Fei Ge; Wei Zhang; Ming Liu] Computer Science Department, Central China Normal University, Wuhan, Hubei, P. R. China;Discipline of ICT, School of Technology, Environments and Design, University of Tasmania, Hobart, TAS, Australia;Computer Science Department, Huazhong Normal University, Wuhan, Hubei, P. R. China;[Xun Gao] Electronic Engineering Department, Wuhan University, Wuhan, Hubei, P. R. China;[Juan Luo] College of Computer Science and Electronic Engineering, Hunan University, Changsha, Hunan, P. R. China
摘要:
The goal of this work is to find appropriate link scheduling schemes to achieve satisfactory end-to-end throughput in wireless multi-hop networks. The algorithm of finding the best path status bitmap is proposed to solve the throughput problem. By analyzing path status, it is found that compressing the path state set can reduce the time complexity. According to this, we describe innovative methods to simplify scheduling of links for long path with large amount of data. Two typical link scheduling schemes with full-duplex radios are proposed, and end-to-end throughput boundary is worked out by analyzing the link capacity and the link active ratio in each scheme. Results illustrate that these schemes may improve end-to-end throughput in wireless multi-hop networks modestly.
期刊:
Security and Communication Networks,2021年2021 ISSN:1939-0114
通讯作者:
Tan, Liansheng
作者机构:
[Huang, Kaijiao; Tan, Liansheng] Cent China Normal Univ, Dept Comp Sci, Wuhan 430079, Peoples R China.;[Tan, Liansheng] Univ Tasmania, Coll Sci & Engn, Hobart, Tas 7001, Australia.;[Peng, Gang] Shenzhen Inst Informat Technol, Sch Comp Sci, Shenzhen 518172, Peoples R China.
通讯机构:
[Tan, Liansheng] C;[Tan, Liansheng] U;Cent China Normal Univ, Dept Comp Sci, Wuhan 430079, Peoples R China.;Univ Tasmania, Coll Sci & Engn, Hobart, Tas 7001, Australia.
摘要:
The Internet is nowadays suffering dramatically serious attacks, with the distributed denial of service (DDoS) attacks being the representative and dominant ones. It is seen that, to stabilize the buffer queue length around a given target under DDoS attacks in the relevant routes is vitally important and helpful to mitigate the attacks and to improve the quality of service (QoS) for normal users. In the current paper, a stochastic queue dynamic model with Le ' vy jump noise, which is affected by the continuous Brownian motion and the discontinuous Poisson process, is worked out to develop a novel and accurate mathematical framework for the stability of a route queue that deals with constant-rate DDoS attacks. This article proposes a security defensive mechanism in the network for solving the network collapse that can possibly be caused by DDoS attacks, otherwise. Particularly, based on the formulation of a stochastic queue dynamic with Le ' vy jump noise, the mechanism that characterizes the behavior of the queue at routers is presented for stabilizing the queue length under constant-rate DDoS attacks. By applying the stochastic control theory into analyzing the performance of queue dynamic under constant-rate DDoS attacks, some explicit conditions are established under which the instantaneous queue length converges to any given target in a route. Simulation results demonstrate the satisfaction of the proposed defense mechanism with sharp contrast to the state of the art active queue management (AQM) schemes.</p>