“Our habits tend to fall into local maxima. We choose well compared to similar alternatives, but ignore options that are totally different yet possibly better. Machine learning algorithms avoid local maxima by occasionally testing random permutations, with techniques like simulated annealing.”