Choosing Algorithm based on Characteristics of the Problem

I believe I first came across a reference to an algorithm that chose an algorithm to solve problems based on characteristics of that problem in a review of Fast Fourier Transform methods. In this article, Leyton-Brown et al. describe a process where they use machine learning to predict the algorithm runtime of various SAT-solvers. Practically, this meant they could write a program that selected the appropriate SAT solver from a pool and, in general, exceed the performance of competitor algorithms overall. The far more useful result, as the paper lightly addresses, would be understanding why SAT solvers do so well with problems that are, complexity-wise, NP-complete. By itself, I am doubtful a machine learning approach will yield a model susceptible to answer "why" but perhaps if the approach incorporated more performance outputs (that collectively help determine the run time) with an augmented machine learning approach as in Schmidt and Lipson?