null pointer error when the ILP solver can't find a solution

The best alignment is null when the solver can't find a solution (or throws an exception), and the feature extraction code doesn't handle this so the code crashes. This seems to happen with some frequency (depending on some kind of random initialization?) on the RTE5 data. Should be a way to recover gracefully from this, e.g. skip this sentence pair and don't use it to update the weight vector.