Ensemble Learning Methodology Research Of Belief Rule Based Systems

With the rapid development of information technology,more and more experts in various fields be attracted attention to the data fusion technology for handle the data with diversity,large capacity,high speed,real-time under complex application background.The methods of data fusion include Bayes probability reasoning,Fuzzy reasoning,and Dempster-Shafter theory.The belief rule base inference methodology using the evidential reasoning is proposed by Yang et al.This method adds a belief frame to the traditional IF-THEN rule,inference and analysis the quantitative information or qualitative knowledge of the input base in the existing belief rule base,and finally provide decision-making basis for decision-makers.However,these available methods of parameter optimization are reasoning for too long and poor portability on account of depending on library functions in MATLAB,or the reasoning ability of the system is affected by the complex formula derivation and only some parameters participate in the training.Apart from this,the existing methods deal with the problem in a single belief rule base(BRB),when the uneven distribution or small amount of training data can lead to the parameter training of the BRB system is not comprehensive and the accuracy of reasoning result is reduced.In order to solve the above problems,this paper proposes a parameter optimization method,an ensemble learning based on accelerating gradient method,and a selective ensemble learning based on multi-objective optimization via Pareto archived evolutionary strategy(PAES).In this paper,the specific research work is as follows:(1)In order to overcome the limitation of the existing method of parameter learning,such as poor portability,difficulty in implementation and long computation time,this paper extended the parametric optimization model and simplify the complex partial derivative process of the gradient descent algorithm(GDA)from the definition of derivation,and proposed the parameters training approach for BRB using the accelerating of GDA.The validity of the proposed method is verified by Himmelblau function fitting and oil pipeline leak detection experiments.(2)The reasoning performance of a single BRB system is degraded when the data distribution is not uniform or the data volume is too small,which leads to the incomplete training of parameters.To solve these problems,this paper proposes BRB-Ensemble system base in GDA via combining the Bootstrap data re-extraction technique and AdaBoost lifting algorithm with BRB system respectively in the regression problem,and the validity of the two BRB-Ensemble methods is validated in the pipeline leak detection and multimodal function fitting experiments respectively.(3)In BBR-Ensemble learning,a large amount of sub-BRB systems are generated due to the difference between subsystems is reduced with the increase of sub-BRB systems.and the prediction speed of BRB-Ensemble system decreased and required more storage space.In order to solve the above problems,this paper proposed the selective ensemble learning of BRB classification system(BRBCS)based via PAES multi-objective optimization in the classification problem,the experimental analysis on the UCI classification data set showed that the method proposed is effective.