Randomized Pruning: Efficiently Calculating Expectations in Large Dynamic Programs