Learning General Policies from Small Examples Without Supervision