Research On Monotonic Fuzzy System Modeling Method

There are a large number of orderly classification problems in real life.In the past years,with the development of research on classification tasks,general classification problems have achieved satisfied classification results.However,these tasks rarely consider the relationship of order,so it may obtain inconsistent decision rules,which requires researchers to study the order relationship between class labels.Takagi-Sugeno-Kang(TSK)fuzzy system has been widely used in regression,classification and decision-making,and has showed a great precision and interpretability.But for the modeling scene with monotonic data,the modeling effect of TSK fuzzy system is not ideal enough.In order to overcome the above challenge,based on the current classic algorithms and the latest research results,this article has conducted the following research work:1.Firstly,a novel monotonically increasing TSK fuzzy system(MCI-TSK)is proposed.The MCI-TSK application exists in the monotonic relationship between response and predictor variables,and is expressed in the form of a monotonic constraint to construct an optimization objective function.The experimental results show that MCI-TSK has better classification performance than the original TSK fuzzy system when dealing with monotonic datasets.2.Secondly,based on the above,the monotonically increasing constraint conditions are extended.By adding a monotonic constraint on the original 0-order TSK fuzzy system model,so as to make the Monotonically 0-order Takagi-Sugeno-Kang(MC-0-TSK)satisfy monotonicity.The MC-0-TSK does not require the monotonic relationship between features and outputs to be consistent,which relaxes the assumption of consistent monotonicity used in most existing methods when dealing with monotonic classification problems.The experimental results show that MC-0-TSK has better classification performance and maintains interpretability compared with the existing methods when dealing with monotonic classification.3.Finally,the prior knowledge of monotonic constraints is further extended to other classifiers.Radial Basis Function neural network is an efficient feed-forward neural network.It has good approximation ability and global optimum characteristic.It has simple structure and good generalization ability.So it has been widely used to the applications of data classification.However,for some special classification scenarios,such as monotonic data scenarios,the RBF neural network does not fully realize its potential.In this paper,a novel monotonic radial basis function neural network model,referred to as Monotonic Radial BasisFunction Neural Network(MC-RBF),is proposed.Experimental results show that the MC-RBF has better classification performance than the original RBF neural network when dealing with monotonous datasets.