You may be interested in this white paper by a Google enginer using a NN to predict power consumption for their data centers with 99.6% accuracy.

http://googleblog.blogspot.com/2014/05/better-data-centers-through-machine.html

Looking at the interals of the model he was able to determine how sensitive the power consumption was to various factors. 3 examples were given for how the new model let them optimize power consumption. I'm a total newbie to ML but this is one of the only examples I've seen where: predictive model -> optimization.

Here's another example you might like from Kaggle cause-effect pairs challenge. The winning model was able to accurately classify whether A->B, or B->A with and AUC of >.8 , which is better than some medical tests. A writeup and code were provided by the top three kagglers.

http://clopinet.com/isabelle/Projects/NIPS2013/