The development of the strategic behavior of peer designed agents⋆

As computerized agents have become more and more common, e-commerce has become 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 a 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. However, 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, while also increasing their means of protection against deceiving agents.