Enviar por SMS: Deep variational reinforcement learning for POMDPs