Opponent modeling in games deals with analyzing opponents’ behavior and devising a winning strategy. In this paper we present an approach to model low level behavior of individual agents using Robocup Soccer Simulation 3D environment. In 2D League, the primitive actions of agents such as Kick, Turn and Dash are known and high level behaviors are derived using these low level behaviors. In 3D League, however, the problem is complex as actions are to be inferred by observing the game. Our approach, thus, serves as a middle tier in which we learn agent behavior by means of manual data tagging by an expert and then use the rules generated by the PART algorithm to predict opponent behavior. A parser has been written for extracting data from 3D logfiles, thus making our approach generalized. Experimental results on around 6000 records of 3D league matches show very promising results.