The notion of buggy procedures has played an important role in recent cognitive models of mathematical skills. Some earlier work on student modeling used artificial intelligence methods to automatically construct buggy models of student behavior. An alternate approach, proposed here, draws on insights from the rapidly developing field of machine learning to develop a student modeling system called Automated Cognitive Modeler (ACM). The ACM system begins with a set of overly general rules, which it uses to search a problem space until it arrives at the same answer as the student. ACM then uses the solution path it has discovered to determine positive and negative instances of its initial rules, and employs a discrimination learning mechanism to place additional conditions on these rules. The revised rules will reproduce the solution path without search, and constitute a cognitive model of the student's behavior. ACM's operation in the domain of multi-column subtraction problems is examined, and some system extensions are proposed. Finally, the generality, psychological validity, and practical utility of this approach to student modeling are discussed. (Author/BW)