Commerce in information goods is one of the earliest emerging applications for intelligent agents in commerce. However, the fundamental characteristics of information goods mean that they can and likely will be offered in widely varying configurations. Participating agents will need to deal with uncertainty about both prices and location in multi-dimensional product space. Thus, studying the behavior of learning agents is central to understanding and designing for agent-based information economies. Since uncertainty will exist on both sides of transactions, and interactions between learning agents that are negotiating and transacting with other learning agents may lead to unexpected dynamics, it is important to study two-sided learning.
We presented a simple but powerful model of an information bundling economy with a single producer and multiple consumer agents. We then explored the pricing and purchasing behavior of these agents when articles can be bundled. In this initial exploration, we studied the dynamics of this economy when consumer agents are uninformed about the distribution of article values. We discovered that a reasonable albeit naïve consumer learning strategy can have a profound influence on market behavior -- in this case, a strikingly bad influence.
Our consumer and producer agents were rather naïve in our first learning experiments. This could be viewed as a criticism of our modeling. However, especially early in the development of adaptive agent intelligence, it may well be that agent-based markets are quite vulnerable to odd behavior and dysfunctional dynamics of the sort we observed. Our consumer agents did not recognize the option value of new information, and thus suffered by not undertaking sufficient exploration relative to exploitation. Our producer agent did not initially adapt to the pathological dynamics induced by the consumer agent naïveté, and thus suffered by relying too confidently on its ``perfect" but static knowledge. Although it was fairly easy for us to see what was going wrong, and to modify the producer agent in a simple way that ameliorated much of the problem, our environment is artificially simple and static. In more realistic settings it may be quite difficult for even relatively intelligent agents to adapt to emergent pathologies. Human markets may not be as susceptible because human behavior is less rote and more reflective. The lesson for agent design is to search for strategies that are dynamically robust and adaptive in the face of substantial uncertainty.
We have started to explore how to make our simple mechanism more robust in realistic settings. For example, the search technique employed by the amoeba algorithm is likely to get stuck at local optima. For the uniform h there is a single peak in the profit landscape, but that is not at all general. Several powerful optimization techniques exist for static landscapes with multiple optima. Extensions to these that handle changing, noisy landscapes may lead to robustly adaptive agent learning strategies.
We have an active agenda of continuing work on this topic. For example, we have begun to consider less naive consumer strategies that balance exploration against exploitation. We are considering producer strategies that adapt based on the number of recent subscribers relative to the producer's model of the optimal number of subscribers. Perhaps most challenging -- but essential to a more general understanding of the problem -- is the extension of our work into an economy with multiple producers who are underinformed about each other's competitive strategies as well as about consumer valuations.
In this and earlier work we have found that initial plausible but simple designs of economically-intelligent agents lead to dynamic market interactions that can be surprising and unsuccessful. The value of intelligent agents in electronic commerce will depend on the ability to understand the problems of learning and adaptivity, and to design agents that interact robustly in the presence of substantial uncertainty about both parameters and the strategies of other underinformed agents.