Mechanism Design, Machine Learning, and Pricing Problems

Consider a seller with multiple digital goods or services for sale, such as movies, software, or network services, over which buyers may have complicated preferences. In order to sell these items through an incentive-compatible auction mechanism, this mechanism should have the property that each bidder is offered a set of prices that do not depend on the value of her bid. The problem of designing a revenue-maximizing auction is known in the economics literature as the optimal auction design problem [Myerson 1981]. The classical model for optimal auction design assumes a Bayesian setting in which players’ valuations (types) are drawn from some probability distribution that furthermore is known to the mechanism designer. For example, to sell a single item of fixed marginal cost, one should set the price that maximizes the profit margin per sale times the probability a random person would be willing to buy at that price. However, in complex or non-static environments, these assumptions become unrealistic. In these settings, machine learning can provide a natural approach to the design of near-optimal mechanisms without such strong assumptions or degree of prior knowledge.