Optimization Methodology Research Of Belief Rule Based Systems

Because humans play an irreplaceable role in the decision-making process,so in engineering practice,integrated use of incomplete or inaccurate quantitative information and subjective information provided by experts on modeling decision problems and analysing issues are very important.In order to effectively use various quantitative and qualitative knowledge with uncertainties,Yang et al.proposed a belief rule base inference methodology using the evidential reasoning(RIMER)approach.RIMER is based on D-S evidence theory,decision theory,fuzzy theory and used a belief structure in the traditional IF-THEN rules.RIMER has the ability to model by vague or fuzzy uncertainty,incomplete or non-linear characteristics and probability uncertainties data.The expert systems using the RIMER approach are called belief rule based systems.In order to enhance the performance of belief rule based systems,Yang et al.proposed optimization models for belief rule based systems to train the parameters of systems to obtain accurate parameters.Liu,et al.used a belief structure in antecedent attributes to enable belief rules better describe uncertainty information and proposed a new inference methodology.The extended belief rule based systems have good performance without parameters learning.However,existing studies ignored the efficiency of belief rule based systems,the accuracy and efficiency are not good enough both in parameters learning and extended belief rule based systems.So belief rule base inference methodology still needs optimizing.The work of this research is as follows:(1)For the operating efficiency and optimization capability of parameters learning based on fmincon function of MATLAB optimization toolbox is not good,this paper proposes the new parameters training algorithm based on particle swarm optimization algorithm.By proposing the strategy reassigning particle velocity and restricting particles in the solution space.To some extent,this approach overcomes the premature convergence of particle swarm optimization problem and particle swarm optimization can be used in a constraint optimization problem.Finally validating the effectiveness of this approach through function fitting and pipeline leak detection example.(2)Belief rule based systems especially extended rule based systems need to traverse the entire rule base to calculate the activation weight for each input during inference process.So the efficiency of the system is not perfect,this paper proposes establish index between disorderly stored rules based on BK tree.The number of search rules will be reduced by a pruning strategy during inference process.The reasoning efficiency and ability of belief rule based systems can be enhance by combining key rules.Finally,the experimental verification is carried out using regression data and classification data.(3)For the reasoning ability and generalization ability of large-scale and complex belief rule based systems is not good,this paper presents an approach based on AdaBoost to train multiple systems and then use appropriate method for combining multiple systems.Finally,experimental verifications show ensemble belief rule based systems have better reasoning ability.