Using the EPAM Theory to Guide Cognitive Model Rule Induction

The isomorphism between rule-based systems and decision trees provides an opportunity to use induction methods associated with decision trees and data mining as the basis for the formation of rules in a cognitive architecture. The EPAM theory of learning, an exemplar decision tree method, provides the core processes for a theory of production formation within the ACT-R architecture which is applied to a traditional concept formation task, the 5-4 category task. Rule creation using this method provides a good quantitative fit to the human performance data. This method is impasse driven, provides generalization, and is tolerant to noisy examples. Excess rules, rather than being pruned, are withered away through utility learning. 1. Inducing Rules from Data The goal of this project is to exploit the isomorphism between decision trees and rule-based representations in an exploration of the use of decision-tree induction (learning) methods in general, and the EPAM theory (Elementary Perceiver and Memorizer, Feigenbaum & Simon, 1984) in particular, for developing cognitive models within the ACT-R cognitive architecture (Anderson & Lebiere, 1998). This work applies to task domains that involve a mapping of continuous, categorical, or symbolic features onto discrete categories. Tasks typical of this type of domain include fault diagnosis, categorization tasks, and strategy selection. One exemplar task, the 54 Categorization Task (also known as the Brunswick Faces Task), will be described in detail below to illustrate the application of decision tree learning to a rule-based representation. In the 5-4 Categorization Task, the goal is to use the presented features of the stimulus to decide which of two groups it belongs to. To accomplish this, the cognitive model must determine which features actually are discriminative, and must learn to rely on them to perform the task. The methodology discussed here is inherently supervised: training examples are given with their classifications. However, “supervised” is meant in this narrow technical sense, and the method is also equally applicable to developing models from streams of human performance data, where the supervision may be far less than perfect, and may simply consist of observing performance that includes both environmental cues and observable actions. Rules as used in many cognitive modeling systems consist of a pairing of cues, or context, with an operator, or action to provide the basis for intelligent behavior. These rules are frequently described simply as condition-action pairs, but can also be thought of as condition-operator pairs in the current context. An often asked question of rule-based cognitive models is: Where do rules come from? The variety of mechanisms proposed include many methods for combining existing rules such as chunking (Newell & Rosenbloom, 1983), compilation (e.g., Taatgen & Lee, 2003), and composition, but each of these begs the question of where those rules that are being combined themselves came from. Recently, work has shifted to learning initial rules directly from instructions and encoding instances as specific rules (Taatgen & Lee, 2003). However, this work provides little guidance on the formation of new rules that combine operators with variable cue structures. The current proposal explores the alternative of an always-on production learning mechanism based on the EPAM theory of learning and recognition. EPAM is, at its core, a decision tree learning process. Although there are many other decision tree learning paradigms in current use, particularly in the area of machine learning, EPAM, as a cognitive theory, provides several specific benefits for this project. One advantage is that cognitive methods are often computationally simpler than AI methods due to their need to reduce the computations to those performable by a person in a limited amount of time (Best, 2005). Another advantage is that cognitive methods depend on a higher amount of pruning of search spaces (e.g., see studies of chess experts by Gobet & Simon, 1998), and as a result may scale more effectively. Yet another advantage is that cognitive methods must be extremely adaptive, and must be able to learn from limited data with potentially changing base-rates, making these systems ideal for deployment in changing, uncertain environments (exactly those environments where more traditional AI systems tend to become brittle and founder).