Learning a nonlinear model of a manufacturing process using multilayer connectionist networks

Control of a manufacturing process can be very risky when the process is incompletely understood. The risk of making adjustments can be deceased by building a model of the process and experimenting with changes to the controls of the model rather than to those of the actual process. A connectionist (neural) network learns a nonlinear process model by observing a simulated manufacturing process in operation. The objective is to use the model to estimate the effects of different control strategies, removing the experimentation from the actual process. Previously it was demonstrated that a linear, single-layer connectionist network can learn a model as accurately as a conventional linear regression technique, with the advantage that the network processes data as they are sampled. Here, experiments with a multilayer extension of the network that learns a nonlinear model are presented