James Andrews, Michael Benisch, Alberto Sardinha, Norman M. Sadeh
The Supply Chain Trading Agent Competition (TAC SCM) was designed to explore approaches to dynamic supply chain trading. During the course of each year's competition historical data is logged describing more than 800 games played by different agents from around the world. In this paper, we present analysis that is focused on determining which features of agent behavior, such as average lead time or selling price, tend to differentiate agents that win from those that don't. We begin with a visual inspection of games from one bracket of the 2006 semi-final rounds. Plots from these games allow us to isolate behavioral features which do, in fact, distinguish top performing agents in this bracket. We introduce an information gain based metric that we use to provide a more complete analysis of all the games from the 2006 quarter-final, semi-final and final rounds. The technique involves calculating the amount of information gained about an agent's performance by knowing its value for each of 20 different features. Our analysis helps identify features that differentiated winning agents. In particular we find that, in the final rounds of the 2006 competition, winning agents distinguished themselves by their procurement decisions, rather than their customer bidding decisions. We also discuss how the information gain analysis could be extended by agent developers to identify potential weaknesses in their entry.
Subjects: 7.1 Multi-Agent Systems; 12. Machine Learning and Discovery
Submitted: May 14, 2007