Given the complexity of preferences in markets such as key word advertising it is hard to believe that the de facto standard, decentralized, local, greedy algorithm (advertisers bid for clicks on keywords) is any where close to being optimal for any reasonable objective (welfare, profit, etc.). In this talk we consider the market design problem from a global perspective. We make connections between machine learning theory and market design theory, where machine learning design problems closely mirror game theoretic design problems. We reduce a general theoretical market design problem to a natural machine learning optimization problem. These theoretical results lead to a number of practical answers to advertising market design questions.