Typically, combining the outputs through majority rules gives the best result. Rather than having 99 nets saying "0" and 1 net saying "1", you're likely to have 51 saying "0" and 49 saying "1". I've seen other methods that would classify as "1" if any classifier gave a "1," but these types of methods are probably more dependent on your tolerance for false positives vs false negatives.

Ultimately, I would say that cross-validation is your best friend. Try a few methods with 5- or 10-fold cross-validation and use that which gives the best cross-validated result.

-Kirk