基于筛选机制的L1核学习机分布式训练方法

为降低无线传感器网络中核学习机训练时的数据通信代价和节点计算代价,研究了基于筛选机制的L1正则化核学习机分布式训练方法。提出了一种节点局部训练样本筛选机制,各节点利用筛选出的训练样本,在节点模型对本地训练样本的预测值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,利用增广拉格朗日乘子法求解L1正则化核学习机分布式优化问题,利用交替方向乘子法求解节点本地的L1正则化核学习机的稀疏模型;仅依靠相邻节点间传输稀疏模型的协作方式,进一步优化节点局部模型,直至各节点模型收敛。基于此方法,提出了基于筛选机制的L1正则化核最小平方误差学习机的分布式训练算法。仿真实验验证了该算法在模型预测正确率、模型稀疏率、数据传输量和参与模型训练样本量上的有效性和优势。

A distributed training method for L1 regularized kernel machines based on filtering mechanism

To decrease the amount of data transferred and the computing cost during training a kernel machine in wireless sensor network, a distributed training method for L1-regularized Kernel Minimum Square Error machine based on filtering mechanism was proposed. First, filtering mechanism of samples was presented and used on each node. Second, with consistency constraint on the local model of each node and its local optimal one obtained by exchanging the local model with its all neighbours, the distributed optimization problem of L1-regularized Kernel Minimum Square Error machine was solved by Augmented Lagrange Method of Multipliers, and the local optimization problem of L1-regularized Kernel Minimum Square Error machine on each node was solved by Alternating Direction Method of Multipliers. Then, the spares model obtained on each node was transferred to its all neighbor nodes. This process iterates until the local model on each node converges. For carrying out this method,a novel distributed training algorithm for L1-regularized Kernel Minimum Square Error based on filtering of samples was proposed. Simulation results prove the validity of the proposed algorithm in terms of convergence, sparse rate of model, the amount of data transferred and the number of samples used in model training.