How can neural networks be used for control systems?

I have been training different types of NN's recently using the NN toolbox. They have proved very useful at mapping non non-linear relationships between inputs and outputs. But I can't see how this can be fed back to control that system.

Lets say I have the input parameters x1,x2,x3 that are mapped to output y by a trained curve fitting NN. If the y value is identified as being too high, then the inputs must change to reduce y. But given the relationship is complex and non-linear, then presumably my trained NN must be used once again to determine what values must change an by how much. I can't currently understand how to do this. If anyone knows how a trained network can be used in a control system, the information would be greatly appreciated.

Thanks,

Steve