Rule based design of a multilayer perceptron

Multilayer perceptrons (MLP) trained with backpropagation algorithm are popular because they offer certain desirable features. However, the training is time consuming which further limits its use in changing environment. We propose a methodology to partially designs MLP network using formal knowledge-rules and augmenting it with another neural network trained using labeled examples. The approach has the potential of reducing the training time and enables online changes in the user environment. The software developed is being used in a process control system.