无线传感器网络下线性支持向量机分布式协同训练方法研究

Research on the Distributed Training Method for Linear SVM in WSN

针对无线传感器网络中分散在各节点上的训练数据传输到数据融合中心集中训练支持向量机(Support Vector Machine,SVM)时存在的高通信代价和高能量消耗问题,该文研究了仅依靠相邻节点间的相互协作,在网内分布式协同训练线性SVM的方法。首先,在各节点分类器决策变量与集中式分类器决策变量相一致的约束下,对集中式SVM训练问题进行等价分解,然后利用增广拉格朗日乘子法,对分解后的SVM问题进行求解和推导,进而提出基于全局平均一致性的线性SVM分布式训练算法(Average Consensus based Distributed Supported Vector Machine,AC-DSVM);为了降低AC-DSVM算法中全局平均一致性的通信开销,利用相邻节点间的局部平均一致性近似全局平均一致性,提出基于一次全局平均一致性的线性SVM分布式训练算法(Once Average Consensus based Distributed Supported Vector Machine,1-AC-DSVM)。仿真实验结果表明,与已有算法相比,AC-DSVM算法的迭代次数和数据传输量略高,但其能够完全收敛到集中式训练结果;1-AC-DSVM算法具有较好的收敛性,而且在收敛速度和数据传输量上也表现出显著优势。

In Wireless Sensor Network （WSN）, transferring all training samples distributed across different nodes to a centralized fusion center for training Support Vector Machine （SVM） significantly increases the communication overhead and energy consumption. Therefore, this paper studies the distributed training approach for linear SVM through the collaboration of neighboring nodes within the networks. First, the centralized linear SVM problem is cast as the solution of coupled decentralized convex optimization sub-problems with consensus constraints on the classifier parameters. Second, the distributed linear SVM problem is solved and derived using the augmented Lagrange multipliers method, and a novel distributed training algorithm, called Average Consensus based Distributed Supported Vector Machine （AC-DSVM）, is proposed. To decrease the communication overhead of global average consensus, an improved distributed training algorithm,named Once Average Consensus based Distributed Supported Vector Machine （1-AC-DSVM）, is presented, which is only based on once global average consensus. Simulation results show that compared with existing algorithms, AC-DSVM has slightly higher iterations and data traffic, but can converge to the centralized training results; 1-AC-DSVM not only has better convergence, but also has remarkable advantage in convergence speed and data traffic.