I have a multi-stage manufacturing process (let’s call them P1, P2, P3 and P4). Each batch has been at these stages for the corresponding times tP1, tP2, tP3 and tP4. My current neural network model predicts the time=0 at P1 correctly, but not for P2-P4. I would like to generate a model that allows me to predict the output values for a batch that has undergone the processes P1-P4 in half of the current times. How do I go about doing this? I don't think this is an extrapolation as I already have data for time points between 0 and the maximum time.
The Science of Manufacturing is very interesting, and often gives very unituitive results. For instance, the Pollaczek-Khinchine equation suggests that the coefficients of variation of the service times will have an effect on the cycle time.
Depending on your current operations, cutting process times in half could make the arrival rate of batches rise above the service rate of a process, and the queue could grow to infinity. I doubt that a reasonable neural network model could accurately predict that.
I think the state of the art is nonlinear constrained simulation and global optimization.