Searching a continuous space of real-valued connection strengths is superior to searching a discrete space of symbolic rules once you move to real, messy data with exceptions and probabilistic truths, because discrete rules keep getting violated whereas real-valued feature interactions handle exceptions gracefully.