Research On Decentralized Ensemble Learning Based On Sample Exchange Among Multiple Agents

In recent years,thanks to advent of time of the Internet,the information era has arrived in.Along with the ongoing maturation and widespread application of cloud computing and artificial intelligence technologies,information technology has gradually converted to the era of the Internet of Things(Io T).The lifestyle of human society has ushered in a brand-new change and has accessed a smart society where everything is connected and enabled to sense.Ordinary physical equipment is not just a "physical machine without emotion",but has similar perception and thoughts compared with humans.At the same time,by 2025,the global data volume is expected to surge into 175 zettabytes(ZB)of data,resulting from unprecedented amounts of data collected by Io T devices though 5G high-speed transmission.Therefore,Data integration and mining become the most important issue that need us to figure out,as well as playing an important role in the development of society.Ensemble learning amounts to a powerful technique that trains and combines multiple weak base learners in the hope of improving the overall performance and having better generalization capabilities than weak-base learners.Existing ensemble learning algorithms cannot deal with a large amount of data efficiently,such as Random Forest,where each base learner's model is based on the whole dataset to sample and train.Moreover,they could result in additional computational costs,and may be impractical.In order to achieve higher efficiency,we propose a decentralized framework,specifically by leveraging distributed edge computing to facilitate ensemble learning techniques.In our framework,network edge nodes(learners)are utilized to process,analyze and build the model for classification and prediction.For the proposed framework,a prior question,which needs to urgently address,is how to balance the need to restrict access of each base learner for a small sub-dataset and achieve high accuracy at the moment.Based on the idea of ensemble learning,this paper takes advantage of the model of edge computing system to design an efficient(low computing costs)and accurate method.In this paper,our framework utilizes the collaboration and autonomy of software agents on data sources to select edge nodes(learners)to help solve complex problems,and at the same time takes advantage of an ensemble approach to train the model for multiple learners and then combines them together.In other words,this paper benefits from a multi-agent system and an ensemble learning approach,in which each base learner in the ensemble approach is treated as an agent and multiple base learners form a multiagent system.Therefore,our proposed Multi-agent Ensemble System(MAE)framework has great advantages on reducing the computational cost and obtaining a reliable efficiency of ensemble learning.In detail,data are distributed among multiple base learners,who exchange their respective data to improve the collective predictive abilities by the specific interaction mechanism in the framework.For a better show of the dominance of our framework,it is evaluated on 20 real-world datasets against several well-known existing ensemble learners.Experimental results show that MAE greatly reduces calculation costs by processing datasets in a decentralized way and obtains improved accuracy scores through sample exchange.The proposed method achieves competitive performance with state-of-the-art ensemble methods whereas the base learners store only a small fraction of the samples,thus leading to significant reduction on computation costs.The basic mechanism shows here.Firstly,the ensemble system framework distributes the work of model training and testing to several independent edge nodes(agents).At the time of initialization,the input dataset is divided into several disjoint clusters by clustering method,and the samples in the clusters are allocated to the corresponding agents.In the progress of training data,agents are allowed to exchange some samples of the local dataset with others.The initial sample of each agent may be unevenly distributed to guarantee the diversity of the agent.However,as the interaction continues,the agent will improve the model from the interaction process to ensure the accuracy of the agent.After the model training is completed by using the learning algorithm,in the prediction stage,the models,locating in peripheral positions(edge nodes),will vote to produce prediction results.MAE algorithm has the following advantages:1)Fast and high performance: Existing ensemble learning algorithms based on sampling the entire data set or training models with high iterations will need more computing costs,while the distributed and incremental access/process to data sources by the MAE algorithm can greatly reduce the computing cost.It can better handle larger sample space and has faster calculation speed than traditional methods.2)Diversity and interaction: At present,the base learner training models of ensemble learning algorithms are limited to local data sets,and the learning method is unitary.By contrast,there is an information interaction method among all agents in our MAE algorithm,which aims to improve the accuracy and diversity of base learners to enhance ensemble learning capabilities.3)Flexibility and scalability: The proposed MAE algorithm decouples important ensemble conditions from the method and is divided into a diversity enhancement phase,an accuracy enhancement phase,and an ensemble combination phase.The method of each stage can be easily adjusted to meet different computing needs.The algorithm is easy to extend and is of strong flexibility as well.