The CMUnited-98 simulator team became the 1998 RoboCup [4] simulator league champion by winning all 8 of its games, outscoring opponents by a total of 66-0. CMUnited-98 builds upon the successful CMUnited-97 implementation [8], but also improves upon it in many ways.
The most notable improvements are the individual agent skills and the strategic agent positioning in anticipation of passes from teammates. While the success of CMUnited-98 also depended on our previous research innovations including layered learning [9], a flexible teamwork structure [10], and a novel communication paradigm [10], these techniques are all described elsewhere. The purpose of this article is to clearly and fully describe the low-level CMUnited-98 agent architecture as well as the key improvements over the previous implementation.
Coupled with the publicly-available CMUnited-98 source code [11], this article is designed to help researchers involved in the RoboCup software challenge [5] build upon our success. Throughout the article, we assume that the reader is familiar with the soccer server [1].
The rest of the article is organized as follows. Section 2 gives an overview of the entire agent architecture. Section 3 describes the agents' method of keeping an accurate and precise world model. Section 4 details the low-level skills available to the agents. Section 5 presents the CMUnited-98 collaborative coordination mechanisms. Section 6 summarizes the RoboCup-98 results and Section 7 concludes.