Computational Argumentation Based Multi-Agent Joint Learning

With the arrival of big data era,distributed data mining has gradually become the mainstream technology in the field of data mining,becoming one of the basic methods to solve the data mining task in large-scale data environment.However,there are still many technical problems in distributed data mining that need to be solved urgently.Among them,the problem of knowledge integration of distributed data mining is related to the quality of the system's global model,which has drawn extensive attention in the field of distributed data mining.Therefore,how to design an effective knowledge evaluation and integration method to solve the task of knowledge integration in distributed data mining has become one of the frontier topics full of research value and challenge.This dissertation focuses on the problem of knowledge integration of distributed classification rules mining.On the basis of multi-agent system technology,this dissertation comprehensively uses argumentation technology,association rule mining technology,sampling technology and reinforcement learning technology,to systematically and deeply study the argumentation model,argument construction and its performance analysis for computational argumentation based multi-agent joint learning method,as well as the optimization method facing large-scale data and dynamic data.The main research work includes the following parts.(1)In order to solve the problem of knowledge integration in distributed data mining,this dissertation first puts forward the idea of learning from argumentation and proves that the argumentation has the learning ability through theoretical analysis,case description and experimental verification.Furthermore,a novel method of joint learning based on argumentation is proposed by combining distributed data mining with argumentation technology.This method is oriented to the knowledge integration task and covers two main functions of the global knowledge extraction and optimization,thus realizing the effective integration of the distributed local knowledge and the optimization of the global knowledge in the application stage.On this basis,using the argumentation model Arena,a multi-agent joint learning model AMAJL is proposed.Then,the basic function and structure of AMAJL is formally defined at the three different levels,which are local knowledge generation,global knowledge generation and global knowledge application and optimization.Finally,we investigate important properties owned by the multi-agent joint learning model AMAJL.(2)In order to show that the argumentation based multi-agent joint learning method is feasible and effective,this dissertation combines the association rule mining technique and instantiates the argumentation based multi-agent joint learning model AMAJL,and proposes an association rule based multi-agent joint learning system ArgAR.The method for instantiating local rules and empirical arguments in ArgAR system is investigated in depth.The algorithm for constructing empirical argument based on association rules,and the main-control flow and algorithm for multi-agent joint learning system,are described in details.Finally,the convergence of the global knowledge extraction process in the ArgAR system is demonstrated by a large number of classification experiments on the UCI public datasets.It is also shown that the high quality global knowledge of the ArgAR system can be effectively integrated and extracted.(3)The sampling technique and multi-agent argumentation technology are combined in this paper,to put forward the concrete application method which combines sampling technique for the large-scale data analysis task,in the method of multi-agent joint learning based on the argumentation.The advantage of knowledge integrationon small-scale samples using multi-agent joint learning based argumentation is analyzed theoretically and in depth.On this basis,it is further verified that the method of multi-agent joint learning based on argumentation can extract high-quality global knowledge from the smaller sample data,through a large number of experimental analysis.Therefore,this method can effectively deal with the knowledge integration task for large-scale data.In the face of the task of large-scale data analysis,this paper applies sampling technique to the multi-Agent joint learning method based on argumentation,and studies and analyzes the effect of the argumentation based multi-agent joint learning method using smaller samples for knowledge integration.Experiments on three datasets show that no matter which sampling strategy is adopted,the classification accuracy of the model with more than 50% sample ratio has difference of only about 5% compared to a model with a 90% sample ratio.It is further shown that the argumentation based multi-agent joint learning method can extract high-quality global knowledge from the smaller sample data and effectively deal with the knowledge integration task in large-scale data.(4)For the knowledge integration task in dynamic data scene,this paper proposes a reinforcement learning based multi-agent joint learning model ArgRL above the method of argumentation based multi-agent joint learning,in order to realize the evaluation and optimization of global knowledge in the classification for dynamic data.This method uses ?-greedy strategy to select actions to classify dynamic data,and uses the Monte Carlo method of reinforcement learning to update and optimize the global knowledge base during the application.Through classification experiments on many public datasets,we show that it is feasible and effective to use the reinforcement learning based multi-agent joint learning model to extract and integrate global knowledge in dynamic data environment.