A novel RL objective maximizing probabilities for recurrent event data is proposed, achieving lower variance and faster convergence compared to traditional methods.