Research On Distributed Learning Methods For Kernel Machines Over Wireless Sensor Networks

Wireless Sensor Network(WSN)is an emerging technology of information collecting and processing,and has been widely used in many fields and has broad application prospects.Classification and regression are the most fundamental and important tasks in various applications of WSN.In WSN,training examples are scattered across different sensor nodes,if all training examples are transmitted to the fusion center by multi-hop routing,a more accurate classifier or regression machine can be learned by the batch learning method with all training examples.However,this centralized learning method has very high communication costs and energy consumption,and is liable to cause congestion and failure on nodes near the central fusion center.As a result,it will lead to the energy imbalance among nodes,and hence greatly reduce the lifetime of the WSN.To avoid and solve these problems,this thesis investigated the distributed learning methods for kernel machines,which based on the distributed optimization theory,1-norm regularization and Markov chain,and depend on in-network processing through collaboration between single-hop neighboring nodes.This thesis has done in-depth research from the following aspects: the decomposition of the optimization problems,the solving of distributed optimization problems,the collaborative approach between neighboring nodes,the construction of the incremental learning problems,the node selection strategy based on Markov chain,and the energy balanced node selection strategy.The major contributions of this dissertation are stated as follows:(1)For the linear kernel models,the linear Support Vector Machine(SVM)was studied from the decomposition strategy,the solving of distributed SVM and the collaborative approach between neighboring nodes,and an average consensus based distributed SVM was proposed.To decrease the communication overhead of global average consensus,an improved once average consensus based distributed SVM was presented.Simulation results proved that the once average consensus based distributed SVM not only has better convergence,but also has remarkable advantage in the convergence speed and the amount of data transmission,so it is considered a rapid and low energy consumption learning method for linear kernel machines.(2)For the nonlinear kernel models,the nonlinear Kernel Minimum Mean Square Error(KM2SE)that incorporates 1-norm regularization was investigated from the decomposition strategy,the solving of distributed optimization problems and the collaborative approach between neighboring nodes,and a model average consensus based distributed sparse KM2 SE was proposed.To decrease the communication overhead,a filtering mechanism of samples was presented and a distributed sparse KM2 SE based on filtering mechanism was proposed.Simulation results proved that the proposed algorithm can obtain almost the same prediction accuracy as that obtained by the batch learning method,and is significantly superior in terms of the sparse rate of model,the communication cost,and the iterations.So it is considered a feasible learning method for nonlinear kernel machines.(3)For the high computation cost of the distributed sparse KM2 SE,an incremental learning algorithm for the sparse KM2 SE was proposed.To adapt to the dynamic change of network topology,the node selection strategy based on Markov chain was studied,and an improved Markov chain based distributed incremental learning algorithm for the sparse KM2 SE was presented.Simulation results proved that the proposed incremental algorithm can obtain almost the same prediction accuracy as that obtained by the batch learning method,and can decrease the computation cost and the requirement for the memory size.The proposed distributed incremental learning algorithm can significantly decrease the computation cost and the amount of data transmission,and can adapt the dynamic change of network topology.(4)For the imbalanced energy consumption caused by the distributed incremental learning algorithm for the sparse KM2 SE,an energy balanced node selection strategy was proposed,and an energy balanced distributed incremental learning algorithm for the sparse KM2 SE was presented.Simulation results proved that it can obtain pretty consistent prediction accuracy with the batch learning method,and can achieve a very simple model.Meanwhile,it has significant advantages with respect to the communication costs,the iteration and the computation costs.Moreove,it can reduce and balance the energy consumption of nodes.Experiments conducted on the test platform further verified the advantages of the proposed algorithms with respect to the transmission energy consumption.And the energy consumption characteristics of sending data and receiving data on the test platform was summarized,it provides a reference for future research.