Abstract
This paper proposes a novel radial basis function network (RBFN) for chemical taste sensor and a method for prediction of optimal substance-concentration of taste. The taste sensor consists of an array of electrochemical sensors and neural networks for gustatory simulation. Our goal is the implementation of a neural-network-based gustatory system that employs the following two functions: (1) estimation of taste from electrochemical sensor responses; (2) prediction of optimal substance-concentration of taste. To implement the former function, we propose a modified version of two-phase optimized RBFN, which has previously proposed by authors, by the adjustment of widths of kernel functions with Parzen window. Furthermore, a new network inversion is proposed for the implementation of the latter function. In the proposed network inversion, the finding rules of the initial input vectors are optimized for modified RBFN. From the experimental results, the estimation accuracy of the proposed RBFN obtained 5.6% higher in average than that of the conventional method. Moreover, the prediction accuracy of the proposed network inversion obtained 10.7% higher in average than that of the conventional method.
