A Study On Event-triggered Distributed Learning Algorithm

In the 21 st century,with the rapid development of the Internet,the amount of data from all walks of life is surging,and the era of big data has come.Nowadays,one of the common concerns of academia and industry is the big data problem with the rapid development of data acquisition and processing technology.Complex,high-dimensional and changeable big data makes it a key and difficult point to seek a fast and efficient data processing scheme.In recent years,the way to process data with parallel technology has become a hot spot,but parallel technology is facing great challenges in the situation that it is unable to collect data centrally and the amount of single computer computing is limited.Data processing in a distributed environment is a better way.Distributed learning is to train the sub datasets distributed on each node of complex peer-to-peer network,so as to achieve the effect of centralized training of global data.Distributed learning requires that all nodes in the network coordinate to reach an agreement.In this process,frequent communication between nodes will lead to high energy consumption and information loss due to random error.Therefore,it is necessary to study the problem of reducing the information traffic of distributed learning algorithm.The purpose of this paper is to use the event-triggered communication mechanism to improve the distributed learning algorithm,and reduce the amount of communication by reducing unnecessary communication.In this paper,stochastic configuration networks(SCN)is used to train and learn datasets distributed on complex peer-to-peer networks to solve the output weight problem of neural networks.SCN is a new supervised learning method in random neural networks.Firstly,the event-triggered communication mechanism is introduced into the distributed learning algorithm based on zero-gradient-sum(ZGS)strategy,and the improved distributed learning algorithm based on ZGS strategy(ZGS-SCN-ET)is obtained.The algorithm uses event-triggered communication mechanism to control the information transmission between nodes.By designing the trigger function,it continuously monitors the state parameters.Only when the trigger error exceeds the threshold and meets the conditions of the trigger function,the node transmits the parameter information to its neighbors and updates its own parameter information in time.Through this way of communication,the improved distributed algorithm based on ZGS strategy is derived.Finally,the simulation results obtained from two datasets prove that the proposed algorithm,ZGS-SCN-ET,can effectively reduce the communication volume of distributed learning algorithm,and achieve the purpose of saving communication resources.Secondly,this paper introduces the event-triggered communication mechanism into the distributed learning algorithm based on the alternating direction method of multipliers(ADMM),and obtains the improved distributed learning algorithm based on ADMM(ADMMSCN-ET).The algorithm also uses the event-triggered communication mechanism to prevent the transmission of variable information with small state change,and deduces the improved distributed algorithm based on ADMM.Finally,the simulation results of two datasets show that the algorithm can effectively reduce the communication of distributed learning algorithm.And the comparison of the two algorithms shows that ADMM-SCN-ET is better than ZGS-SCN-ET on the whole.