论文部分内容阅读
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.
Aiming at the problem of high data communication cost in wireless sensor networks (WSN) when training data is transmitted through multiple hops to a data center for centralized training, based on the sparseness of L1 regularization, (KMSE) learning machine in the network.Firstly, under the constraint that the local optimal model between the node model and the neighboring node is consistent with the prediction value of the local training sample, the parallel projection method and the alternation Direction multiplier method is used to solve the optimization problem of L1 regularization KMSE. Then, when each node converges to the local stability model, the average consistency algorithm is used to achieve the global consistency of each node sparse model.Based on this method, (L1-DKMSE-PP) training algorithm for L1 regularization KMSE learning machine based on the parallel projection method.The simulation results show that the L1-DKMSE-PP algorithm can get the equivalent prediction results and the sparse prediction compared with the centralized training algorithm Model, more importantly, can significantly reduce the cost of data communications in nuclear learning machine training.