Improving Learning Performance by Applying Economic Knowledge

摘要

Recent advances in networked information technology have led to the emergence of a digital economy in which information goods are traded online by both human and computational agents. In these environments, a producer of information goods is faced with a potentially daunting problem: how to position itself within a large and dynamic product space. By their nature, digital goods can be easily bundled, unbundled, and re-bundled into different configurations. Rather than selling "a newspaper", a producer can sell the whole news collection, categories of articles, individual articles, graphics, data tables, etc. Low marginal cost and the easy configurability of digital information goods make unbundling and rebundling attractive, allowing a producer to precisely target consumers, at the expense of an exponential proliferation of possible product configurations. Digital economies are also typically highly dynamic. Due to the large product space, a producer will typically need to search for good product configurations. In addition, both the presence of competing producers and nonstationarity in the consumer population can introduce secondary dynamics. This implies that a study of these systems must take into account not only their equilibria, but also potential non-equilibrium behavior while producers are learning. In this paper, we consider the problem of competing information producers positioning their offerings when faced with a heterogeneous consumer population. We focus on the role that economic knowledge can play in reducing the producers' search burden. When producers must simultaneously learn what to offer and how to price it, the naive learning problem becomes very difficult. However, if producers use even a very limited amount of economic knowledge, they can sharply reduce the difficulty of each learning problem, locating a profitable niche and extracting a significant fraction of the equilibrium profit. This illustrates the importance of understanding the economic principles underlying an otherwise difficult multiagent learning problem.