Automated Learning of Pricing and Bundling Strategies in Information Economies

The advent of automated commerce and the ability to electronically distribute information goods, or articles, have created a new set of problems for producers of information content. Since information goods have negligible marginal cost, producers are presented with a huge space of potential bundling and pricing strategies to choose from. Producers potentially have two separate decision problems to solve: what to offer and how to price it. For example, in deciding what to offer, producers may choose to offer articles from a wide variety of categories, or just offer in-depth content in a few selected categories. In deciding what to charge for their goods, a producer might choose from a number of schedules such as a flat fee, a per-article price, or a schedule which changes nonlinearly based on the number of articles purchased. The choice of schedule and goods offered will impact both the producer’s per-iteration profit and also what it is able to learn about consumer preferences. In an online environment with a large number of goods and rapid transaction times, it will be advantageous for a producer to automate this decision-making. In (Brooks et al. 1999), we describe this problem and show how a monopolist producer can acquire higher aggregate profit by learning a simpler pricing schedule. Our current work considers two other factors that complicate the producer’s learning and decision-making problems: the presence of other producers and differences in article content. In a world in which a producer is competing with other producers for a market of consumers, the producer must consider both the strategy of other producers and the preferences of the consumers. Typically, neither of these are known, but instead must be learned over time. What makes this difficult is that the other producers are also trying to learn about consumer preferences, so producer strategies are not necessarily stationary. One question we are concerned with is the extent to which reasoning about opponent strategies is actually useful; there will be a tradeoff between an increase in the degree of modeling detail and the marginal gain in the value of the information acquired. One factor that makes this problem different from a traditional game theory problem is that we are explicitly interested in the nonequilibrium rewards that a producer receives. That is, it’s not enough for a producer to eventually learn an