I wonder if this could be used to automatically suggest new policies? Presumably the optimizer will tell you exactly which input variables to tweak to increase the final number/gpd/which ever. As the saying goes "any metric is bad once you optimize for it", so maybe that would lead to terrible politics. That makes me worry if large actors can use the model to make any self-serving policy suggestion "look" like a good…
Yes, you tweak variables and look at the impulse-response functions. That's what those models are built for.