We propose a network architecture based on adaptive receptive fields and a learning algorithm that combines both supervised learning of centers and unsupervised learning of output layer weights. This algorithm causes each group of radial basis functions to adapt to regions of the clustered input space. Networks produced by this algorithm appear to have better generalization performance on prediction of non-linear input-output mappings than corresponding backpropagation algorithms and requires a fraction of the number of connection weights required by fixed center radial basis. For a test problem of predicting product quality of a reverse osmosis desalination plant, the network learns much faster than a three-layer perceptron trained with back-propagation, but requires additional computational burden.