Abstract— Opponent Modeling is one of the most attractive and practical arenas in Multi Agent System (MAS) for predicting and identifying the future behaviors of opponent. This paper introduces an approach towards opponent modeling in RoboCup Soccer Coach Simulation. In this scene, an autonomous coach agent is able to identify the weaknesses or patterns of the opponent by analyzing the opponent's past games and advising own players. To gain this goal, we introduce a 3-tier learning architecture. At first, by gathering data from the environment, sequential events of the players are identified. Then the weaknesses or patterns of the opponent are predicted using statistical calculations. Eventually, by comparing the opponent patterns with the rest of team's behavior, a model of the opponent is constructed. According to this architecture, coach models the opponent and to simplify pattern recognition, provides an appropriate strategy to play against the opponent. This structure is tested in RoboCup Soccer Coach Simulation and MRLCoach was the champion at Iran Open 2006. Index Terms—CLang, Decision Tree, Opponent Modeling, Pattern, RoboCup. I. INTRODUCTION ULTI AGENT SYSTEM is one of the sub-disciplines of artificial intelligence which was introduced for the purpose of defining the rules and principles for developing complex systems and provides a mechanism for cooperating the agents [1]-[2]. In real-time environments, multi agent systems need agents that are able to act automatically and as a part of a team. Modeling in multi agent system environments predicts the future behaviors of opponent and propose an appropriate counteraction [3]. RoboCup is an MAS environment and opponent modeling plays a crucial role in this context. In this domain, two teams formed by autonomous agents connect to a server and play a simulated football [4]. A coach agent can receive the complete and noiseless information from the field and to enhance the performance sends messages in format of the standard coach language, called CLang, to its players [5]-[6]. 1-4244-0537-8/06/$20.00 ©2006 IEEE Recently for emphasis on opponent modeling, coach competition has had changes, so that coach is in charge of identifying the weaknesses and strengths of the opponent, which are called patterns, from other behaviors of the opponent, called base strategy. The 2006 coach competition rule defines pattern and base strategy as: pattern: A simple behavior that a team performs which is predictable and exploitable for the coaches. base strategy: The general strategy of the test team regardless of the pattern in it. To exemplify this, a pattern may be a sequence of consecutive passes between some particular players, clearing the ball to the outside of penalty area by defenders, or a different formation of players between pattern and base, etc. Our work is focused on opponent modeling and online pattern identification. For this purpose, MRLCoach receives the previous plays of the opponent as two log files of pattern and base, and by analyzing them identifies the events occurred such as pass, shoot, dribble, etc. Then for pattern recognition, we use chi-square test [7]-[8], to analyze the possible relation between an event and a sequence of previously occurred events. The eventual model of the opponent could be a collection of multiple identified patterns. Now, using a radix tree [9], we compare models constructed from the pattern and base log files and store the difference between them as the final model of the opponent in the model repository. Coach makes models for each of the pattern and base log files. In online mode, observing the live game, coach exposes an online model of the opponent and compares it with the stored models in repository. When a model matches this model, coach reports it as the current opponent model to the server. Additionally, to recognition of the model, coach advises the players with a suitable strategy. The remainder of this paper is organized as follows: At first, a 3-tier learning architecture for predicting and exposing the opponent behavior is presented. Afterwards, we explain how this process of learning is accomplished in online game and a proper strategy is suggested against the opponent team. Continuing on, by analyzing the results of the Iran Open 2006 competition, we demonstrate the capabilities of MRLCoach. Finally, we conclude with a brief description of related work and our conclusions. Mining Opponent Behavior: A Champion of RoboCup Coach Competition * Ramin Fathzadeh, * Vahid Mokhtari, * Morteza Mousakhani and *† Fariborz Mahmoudi * Mechatronics Research Laboratory Department of Computer Engineering, Islamic Azad University, Qazvin Branch, Qazvin, Iran {fathzadeh, mokhtari, mousakhani}@mrl.ir † Iran Telecommunication Research Center mahmoudi@itrc.ac.ir M 80