On a Dynamical Analysis of Reinforcement Learning in Games: Emergence of Occam's Razor