Exploiting Structure And Relaxations In Reinforcement Learning And Stochastic Optimal Control