The goal-gradient hypothesis and maze learning.