R help - Aug 2012 - Optimizing a model toward desired outputs once trained?

I didn't get any responses to this question on stats.SE:
-
http://stats.stackexchange.com/questions/34415/optimization-of-models-ann-radial-basis-etc-in-r-to-target-predictor-levels

What I'm looking for, using neuralnet as an example, is how to guide a
model toward an output profile once the model is trained. For example:

model1 <- neuralnet(formula=out1 ~ intput1 + input2 + input3 + input4,
data=train, hidden=6,
                   threshold=0.05, linear.output=TRUE)

model2 <- neuralnet(formula=out2 ~ intput1 + input2 + input3 + input4,
data=train, hidden=6,
                   threshold=0.05, linear.output=TRUE)

And so on.

Once I've trained each of these models and am satisfied with their
prediction of test data... how can I optimize toward a desired blend
of outputs based on these models? The above is simplistic with two. I
model chemical formulations and measure various responses. I may have
10+ measures with corresponding models for each. From there, I would
like to figure out how to predict the likely formulations that would
produce a result with targeted values for each response.

There's a lot of data on training models and using it to predict
outputs for new inputs... but not on how to use a model to provide
suggested input values if you know what you want your outputs to be.

Hopefully this makes sense.


Many thanks,
John