|
Literature shows that excessive structures of systems and unacceptably heavy computation often restrain the applications of neuro-fuzzy systems. Thus, a novel structure simplification algorithm , SSA ,is proposed as the preprocess of all neuro-fuzzy systems.
Two stages of processes comprise SSA. First, by combining the concepts of correlation matrix and Principal Component Analysis, the number of input variables can be significantly reduced without much loss of information content about training data. In order to minimizing the number of memberships used for fuzzy inference, the technique of sliding-window scanning is then applied to efficiently partition the domain space for each pairs of selected input and output signals. After the operation of SSA, the initial status of a neuro-fuzzy system is automatically and properly established without any human expertise.
To illustrate the performance and generality of SSA, some benchmarks were respectively tested in both the well-known Adaptive Neuro-Fuzzy Inference Network (ANFIS) and the hybrid system of SSA and ANFIS by using MATLAB simulator. Those benchmarks include the approximation of a sinc function with two inputs, the prediction of Mackey-Glass chaotic time series with four inputs, and the prediction of Gas furnace time series with ten inputs. Experiments showed that, in problems of both function approximation and prediction, SSA can significantly reduce the nodes used in ANFIS, speed up the learning and inference processes, and largely improve the capabilities of identification and generalization.
|