Within a few years,
we anticipate that software agents will participate in a wide variety
of commercial transactions, and may even become economic players
in their own right [8]. One important
domain for agent economies is the production and distribution of
information goods and services, such as news articles, entertainment
and other service reviews, and instructional materials. As is
the case with physical goods, producers can reduce costs
or increase the surplus extracted from consumers by
bundling information goods together 
.
We expect negotiations over the composition and prices for
bundles to become a natural application for
software agents. Agents representing producers could use
up-to-date information about market conditions and consumer demand
to explore not just prices, but also locations in a
multi-dimensional bundled product space. Agents representing
consumers could
gather relevant information by purchasing bundles
from multiple producers, always keeping an eye out for new
opportunities that arise as the producers change their
offerings in an effort to gain a competitive advantage.
While they will have their limitations, software agents
will have important advantages over
humans: they are likely to respond more quickly to changing
market conditions and deal more easily with pricing structures
that are more complex than might be feasible
in an all-human economy.
On the other hand, economic software agents are
endowed with less common sense than their human counterparts,
and interactions among them may lead to strange and
undesirable market dynamics . To avoid adverse and possibly disastrous individual, firm and market
consequences, humans must design their economic software agents carefully,
taking into account their likely interactions with their
environment and with other agents.
Although there is a growing literature on commodity
bundling in the context of information goods [1, 4],
much remains to be learned before we can design
competent software agents that buy and sell bundled information
goods. Very little is understood about strategic search over
the joint information good price and product space (that is, search
for which products to offer, in which combinations, at what
prices). Agents
competing in this space must learn about the distribution of
preferences across a heterogeneous and changing customer population,
must learn about the strategies being followed by their competitors,
and then must optimize their strategies to take into
account the new understanding of customers and competitors.
Consumers, on the other hand, attempt to optimize their purchases
by learning about the quantity, quality
and price of bundled items offered by the various producers, all of which
may change over time.
Learning and optimization are both difficult search problems, and in
the market context they are closely related.
Our recent work on strategic pricing for bundles [5] represents the first study of bundling in a market with multiple producers. In that work, we assumed that all of the relevant parameters were known by the producers. We distinguished between conditions in which firms prefer to offer comprehensive bundles at a single price, those in which they prefer to sell items individually, and those in which they prefer to offer consumers a choice between a bundle or individual components.
In this paper we examine a considerably more general problem, although for this initial foray we maintain some simplifying restrictions. We envision an agent economy with one information broker agent and many consumer agents. The consumer agents are heterogeneous: they value each item differently, and these values are drawn from a different distribution for each agent. To set profit-maximizing prices, the broker agent desires, but may not know, the parameters of the consumers' valuation distributions. The consumers also may not know their valuation parameters until they gain experience with the broker's offered information goods. Therefore, both sides wish to learn. The broker can set prices strategically, learning about consumer valuations from their purchasing behavior at different prices. The consumers can purchase strategically, learning about the distribution of information good values by sampling.
This environment is very rich and will permit us to explore many interesting questions. In this paper we focus on one surprising result: that when consumer agents follow a plausible but overly naïve learning strategy, even if the producer is fully informed (but also somewhat naïve), the economy can continuously degenerate with disastrous overall performance. We find a simple explanation for the initial failure, and develop a simple improvement to the producer agent's strategy that largely ameliorates the problem. But in the process we learn an important lesson: dynamic market interactions when there is substantial uncertainty can lead to pathological outcomes if agents are designed with ``reasonable'' but not sufficiently adaptive strategies. Thus, in programmed agent environments it may be essential to dramatically increase our understanding of adaptivity and learning if we want to obtain good aggregate outcomes.
This paper is the first in a series of studies directed towards understanding and developing robust adaptive learning techniques that are effective for individual agents and lead to acceptable collective (market) behavior. We are also extending the model to study the much more common setting in which there is competition among multiple producers of information goods. Of course, this setting exacerbates the learning problem, since the producers need to learn about each other's strategies in addition to learning about consumer valuations.
In the next section we discuss the model. We introduce two relevant distributions: first, a parameterized distribution g from which a given consumer's valuations are drawn, and second, a distribution h describing the population of consumers -- the distribution from which an individual consumer's valuation distribution parameters are drawn. For tractability, we assume that the broker is restricted to two-part tariff pricing schemes [16]: it can charge consumer agents a subscription price to examine the current items, and a per item price for each item the consumer subsequently purchases.
In a general analysis section, we derive the optimal strategies for fully informed broker and consumer agents as functions of the article value and consumer type distributions g and h. Then in Section 4 we explore specific cases of exploration and exploitation through analysis and simulation. When agents are underinformed and engage in learning, we consider plausible (but not necessarily fully optimal) agent strategies, and examine the resulting system outcomes. We close with a summary of our current results and plans for future work.