Précis: Bargaining Automation Project


If we knew how to build computer programs that could negotiate for us effectively and reliably, the benefits of having such knowledge (and the resulting artificial agents, or AAs) would be profound.

The possibility of building effective AAs for bargaining and negotiation raises the prospect of greatly reducing transaction costs as well as improving transaction outcomes with electronic markets, both for business-to-business electronic commerce and for consumer-based electronic shopping. There are three principal reasons for this.

  1. With regard to costs, not only might transactions be effected much more efficiently through automation, but substantial portions of the inevitable decision making that leads to transactions might also be substantially automated. Artificial agents might be made to substitute for human effort, thereby yielding additional benefits in cost and time through the use of electronic markets.
  2. Regarding outcomes, it is well known that humans do not negotiate well, in that they often "leave money on the table" by coming to agreements that can be improved for all involved. The prospect for AAs is that they could (in many circumstances) overcome temporal pressures, cognitive limitations, fatigue, and conflicts of interest that reduce the effectiveness of human negotiators, even in the most routine and simple forms of commercial negotiation.
  3. Effective and efficient systems for negotiation would lead to more reliance on negotiation (including auctions) and less on posted prices. A trend in this direction is already evident in the market for airline transportation. Posted prices are so variable that airlines are, in effect, close to conducting ongoing auctions for their services. The benefit of such a regime is that permits a more rapid and finely-tuned response to market conditions, resulting in a general improvement in efficiency.

While the benefits of having effective AAs for negotiation are substantial, so are the challenges to be met before this can be realized. Existing theory is simply inadequate for the purposes of telling us what strategies the AAs should employ or how these strategies will perform. Recent results, from a number of sources, including the principals in this project, auger well for the possibility of building effective bargaining AAs through use of machine learning techniques. Quite a number of experimental agents have been built, for a fairly wide-ranging set of bargaining contexts, and have (via machine learning) acquired strategies that perform well compared not only to humans, but also to that prescribed for perfectly rational agents by the mathematical theory of games. Still, much remains to be learned.

The purpose of the Bargaining Automation Project is to develop the fundamental knowledge--in the form of concepts, theory, and techniques--needed for building artificial agents that will be able to negotiate effectively in electronic commerce. To date, we have achieved considerable success this objective with AAs in stylized settings. We are in process of broadening and deepening this line of research.

An intermediate-term goal of the Bargaining Automation Project is to develop a software system, called the Bargaining Automation Laboratory, that would facilitate multiagent negotiation experiments on realistic problems bearing close resemblance to those encountered in electronic commerce. Our objective the Bargaining Automation Laboratory is to facilitate experiments among AAs, among human subjects, and among mixtures of AAs and humans. In addition, the Bargaining Automation Laboratory would provide an environment for investigating user interfaces and other aspect for the support of human negotation in electronic commerce.

Steven O. Kimbrough and James D. Laing | University of Pennsylvania | 3620 Locust Walk | Philadelphia, PA 19104-6366 | 215-898-5133 and 215-898-1175 | kimbrough@wharton.upenn.edu and laing@wharton.upenn.edu | This page: http://opim.wharton.upenn.edu/~sok/comprat/bapprecis.html | Created: Febraury 4, 1996 | Last revised: February 4, 1996