R-LAtte: Visual Control via Deep Reinforcement Learning with Attention Network