Improve, search, confirm and hand over

Now we make the model better, fairly. You'll look at where it fails, add features one change at a time, see how a model can memorize instead of learn, and use cross-validation to tell real gains from luck. Then the big question: if we keep trying things, won't something look good by chance? Tuning turns out to be a search problem, which leads to formulating optimization problems and to genetic algorithms, the method behind Earth. Finally we open the test set once, write a handover note, and take a short tour of the wider field: unsupervised learning and deep learning.