The problem of applying machine learning techniques to search problems (I mean "search" in the traditional AI sense that includes theorem proving and planning, not in the Google sense). To me this is the most exciting area of AI research, and it's still very much wide open.