Understanding how people design trading agents over time

As computerized agents are becoming more and more common, e-commerce becomes a major candidate for incorporation of automated agents. Thus, it is vital to understand how people design agents for online markets and how their design changes over time. This, in turn, will enable better design of agents for these environments. We focus on the design of trading agents for bilateral negotiations with unenforceable agreements. In order to simulate this environment we conducted an experiment with human subjects who were asked to design agents for a resource allocation game. The subjects' agents participated in several tournaments against each other and were given the opportunity to improve their agents based on their performance in previous tournaments. Our results show that, indeed, most subjects modified their agents' strategic behavior with the prospect of improving the performance of their agents, yet their average score significantly decreased throughout the tournaments and became closer to the equilibrium agents' score. In particular, the subjects modified their agents to break more agreements throughout the tournaments. In addition, the subjects increased their means of protection against deceiving agents. Copyright © 2008, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved.