A Cognitively-Oriented Architecture Confronts Hard Problems

This paper describes a cognitively-oriented architecture that facilitates the development of expertise. Based on knowledge about human decision making, it integrates multiple representations, multiple decision-making rationales, and multiple learning methods to support the construction of intelligent systems. A constraint solver implemented within the architecture engineers a problem-solving paradigm. This program manages a variety of search heuristics and learns new ones. It can transfer what it learns on simple problems to solve more difficult ones, and can readily export its knowledge to ordinary solvers. It is intended both as a learner and as a test bed for the constraint community. Both the program and the architecture are ambitious, ongoing research projects to support human reasoning. The thesis of this work is that AI architectures for learning and problem solving can benefit from what works for people. Since people function remarkably well in a complex world, other problem solvers may profit both from descriptions of human problem solving and from the devices that people rely upon to achieve their goals. This paper reports on a cognitively-oriented program that learns to solve difficult constraint problems, has rediscovered an important result in graph coloring, and has learned new heuristics that readily export to improve ordinary CSP solvers (Epstein et al., 2002; Epstein & Freuder, 2001). The program can determine when it has finished learning, and can learn when and how to modify its own reasoning structure without sacrificing performance. Here, an architecture is a system shell within which an application program for a particular domain (set of related problem classes) can be constructed. To produce a program, one applies an architecture to a domain by providing it with knowledge: what to learn, how and when to learn it, and how to reason in the domain. An architecture is cognitively-oriented if it simulates significant characteristics of human problem solving.. 1 Since a human expert solves certain problems faster and better than other people (D'Andrade, 1990), programs produced by a cognitively-oriented architecture that facilitates the development of expertise should eventually solve faster and better than others too. Traditionally, an architecture engineers a domain-specific program. The application described here, however, formulates knowledge about constraint satisfaction. Thus it engineers a problem-solving paradigm, rather than a single domain. Furthermore, one can specialize the program further, say, for graph coloring Copyright 2004, American Association for Artivicial Intelligence (www.aaai.org). All rights reserved. or scheduling. The next section references some cognitive science results and interprets them for an architecture. For clarity, we initially illustrate these ideas with an autonomous mobile robot moving toward a goal object in a dynamic environment. Subsequent sections describe how those results influence the formulation of a particular cognitively-oriented architecture, and how that architecture supports a program that learns to solve constraint problems. The final section discusses related and current work. Some Contributions from Cognitive Science The results discussed below are particularly relevant in the creation of a cognitively-oriented architecture, and in any attempt to understand or communicate with one.