Linear Reinforcement Learning with Ball Structure Action Space