面向WSN的稀疏核学习机分布式训练方法

A Distributed Training Method for Sparse Kernel Machine over WSN

针对无线传感器网络(WSN)中,经过多跳路由传输训练数据到数据中心进行集中式训练时存在的高数据通信代价问题,基于L1正则化的稀疏特性,研究了仅依靠邻居节点间的协作,在网内分布式协同训练核最小均方差(KMSE)学习机的方法.首先,在节点模型与邻居节点间局部最优模型对本地训练样本预测值相一致的约束下,利用并行投影方法和交替方向乘子法对L1正则化KMSE的优化问题进行稀疏模型求解;然后,当各节点收敛到局部稳定模型时,利用平均一致性算法实现各节点稀疏模型的全局一致.基于此方法,提出了基于并行投影方法的L1正则化KMSE学习机的分布式(L1-DKMSE-PP)训练算法.仿真实验结果表明,L1-DKMSE-PP算法能够得到与集中式训练算法相当的预测效果和比较稀疏的预测模型,更重要的是能显著降低核学习机训练过程中的数据通信代价.

In wireless sensor network（ WSN）,the centralized learning method by transmitting all training samples scattered across different sensor nodes to a centralized data center to train classifier will significantly increase the communication cost. To decrease the communication cost in transmitting training samples,a distributed learning method for kernel minimum squared error（ KMSE） by incorporating L1 regularized term was studied,which just relies on in-network processing between single-hop neighboring nodes. Each node obtains its local optimum sparse model by constructing the optimization problem of L1 regularized KMSE based on its local training samples and solving it using parallel projections and alternating the direction method of multipliers,then a consistent model is achieved on all nodes by using the global average consensus algorithm. For carrying out this method,a new distributed training algorithm for L1-regularized kernel minimum squared error based on parallel projections（ L1-DKMSE-PP） was proposed. Simulations show that L1-DKMSE-PP can obtain almost the same prediction accuracy as that of the centralized counterpart and a sparser model,and more importantly,it can significantly reduce the communication cost.