Backwards Neural Networks Optimisation

The problem of predicting multivariable process inputs for a given set of process outputs is solved by training a combination of partitioned feedforward backpropagation neural networks. Training of a single multi-layer network produces large unacceptable errors in the predictions due to the absence of monotonic process output functions. However, by configuring a system with parallel partitioned networks linked with a single output network, significant improvement in accuracy of the prediction model was achieved. Percentage errors in the prediction were found to be reduced from a maximum of 73\% using a single network system to 38\% using parallel networks linked with a single output network. The improvement in accuracy is largely dependent on the method of partitioning, the training algorithm and filtering of the training data.