OPTIMISATION OF RELATIVE ERROR CRITERIA IN NONLINEAR NEURO-FUZZY IDENTIFICATION

摘要

In this paper an approach for the minimisation of user defined performance criteria in nonlinear Neuro-Fuzzy identification is presented. Neuro-Fuzzy models are an effective means to partition nonlinear functions into subdo-mains which are then described by local regression models. In many practical applications varying noise in measured data is an important problem both for regression model parametrisation and partitioning based on available data. As a solution approach the proposed algorithm allows for the incorporation of relative performance criteria to achieve a desired relative accuracy with a small number of local models. The main advantage of the proposed algorithm is that relative weights are not only used for the computation of the local model parameters but also for the determination of the region of validity of the local models. Using the proposed algorithm the optimisation of the partitions is focused on the regions of interest of the input space regarding a relative (local) performance criterion. The effectiveness of the proposed concepts is demonstrated by means of an illustrative and an application example.