Finding Paths of Least Action with Gradient Descent