Modifying the generalisation characteristics of a neural network with interactive reinforcement training

An interactive reinforcement training approach to modify the generalisation characteristics of a backpropagation neural network is proposed. The objective is to ensure that the network is capable of recognising significant training data even they are low in number. The interactive process will reinforce the important data by duplicating them. It ensures that the significant data are included in the final network. A case study of porosity prediction in petroleum exploration is used to illustrate this approach. Results have shown that the network's generalisation ability is modified to include the important outliners while avoiding overfitting. It is also useful in cases where training data are difficult or expensive to obtain.