Using background knowledge to speed reinforcement learning in physical agents