Discretizing actions for deep Q learning