Symmetry in Markov Decision Processes and its Implications for Single Agent and Multiagent Learning