For those who aren't apriori familiar with it, heres a wee summary. Its been a few years since i've worked in the area, but the facts havent changed in the mean time.
Smoothed analysis asks the question: "If I apply some small random noise on the inputs, whats the average/expected runtime of an algorithm on that noised input?". Then to get the "smoothed complexity" you pick the family of inputs that maximize the average runtime of the algorithm.
Turns out for many algorithms which have funny "exponential time" corner cases, but are otherwise polynomial time in practice (like the simplex algorithm used in solving linear programs), the smoothed complexity is polynomial! So smoothed complexity gives a nice way of explaining "slow in theory, fast in practice".
Its worth noting that smoothed analysis only really makes sense on algorithms where perturbing the input only changes the answer a little bit, so you need some sort of "numerical stability" style property for the underlying algorithm one way or another.