Under-Approximating Expected Total Rewards in POMDPs