Optimisation of radial basis function neural networks using biharmonic spline interpolation

Biharmonic spline interpolation has been applied as an optimisation tool to study response surfaces of bi-directional data. Both regularly and randomly spaced training data yielded results with prediction errors in the range 0.1 to 10%. Practical application of the technique has been demonstrated by optimising both the spread parameter and the number of neurons in the hidden layer of radial basis function (RBF) neural networks. The efficiency and practical application of this optimisation approach is demonstrated by the prediction of the auto-ignition temperature (AIT) values of 232 organic compounds using quantitative structure–property relationships (QSPR) with six descriptors. It is concluded that this optimisation strategy is fast and provides a very flexible way of modelling non-linear systems in general.