I don't get what the difference is between "machine learning" and plain-old constraint solving. From looking at the abstract of his paper, he's solving an NP-Co

I don't get what the difference is between "machine learning" and plain-old constraint solving. From looking at the abstract of his paper, he's solving an NP-Complete problem using a new approximation algorithm. So, the real contributions of his work are 1) mapping his particular sensor network problem to the NP-Complete complexity class, and 2) an improved approximation algorithm. The newspaper article seems astounded that multiple (real-world) problems can be mapped to the same algorithm.