Computation and Human Experience

During the 1980s, there were two main approaches to the computational study of human intelligence. The first, and largest, was the symbolic approach, derived from the work of Church and Turing, and later championed by two giants in the field, Simon and Newell. These researchers formulated the Physical Symbol System Hypothesis, according to which all of human intelligence can be expressed as a process of search in a symbolic state space transformed through the use of equally discrete operators. The other camp, mostly hiding in the shadows for much of this time, derived from control theory and the servomechanisms of WWII. They held that the human brain was not a discrete symbol-processing entity but rather something constantly in direct contact with a continuous world. Although this group found its closest computational champions in Rosenblatt and Rumelhart, it paled in comparison to the promises and research invested early on in the symbolic approach. Agre's book, Computation and Human Experience, was written as a call to arms for researchers in the symbolic tradition, a challenge to critically re-evaluate their own ideas and methods. In contrast to the "mentalist" juggernaut, Agre proposes an interactionist view of cognition, and shows how such an approach can be reconciled with the technical practice of constructing computational models. The book achieves a rare balance of philosophical argument with computational theory, though in both sides experts will be able to find holes in Agre's arguments. However, the biggest problem with this book is its relevance to the current state of affairs. Much of this work is based on the research that went into Agre's doctoral dissertation (completed in 1988), and in the 2 decades that have passed the situation in cognitive science has improved dramatically. Embodied cognition is not a dark art but an accepted and thriving practice, deictic representation is more commonplace, and even "rule-based" production system architectures like Anderson's ACT-R have found the representations and techniques necessary to interact with a dynamic and continuous world. Whether or not Agre's book has contributed to the current and improving state of affairs is a matter for speculation (my feeling is that it has), but it is most important that no reader today mistake this book's perspective ("situated" in the mid-1980s) as representative of the current status of cognitive science.