Others have done something similar, for example see: http://satirist.org/learn-game/systems/neurochess.html. It is not a new idea that you can learn an evaluation function this way. But in both these cases the resulting performance is not that good. A neural network may not be the best runtime model, plus these experiments are not starting with a good, fast search engine. Better results have been obtained by parameter tuning: in most cases not learning from scratch, but trying to improve an existing evaluation.