Methodologic Assumptions

The architects of TETON argue that the following capabilities must be supported in any architecture proposed as a unified theory of cognition in order to explain the human capabilities of situated action and rapid recovery from interruption when goal memory and short term memory are limited. These memories are purported to overwrite themselves when overextended such as in the case of extended thought or interruption.

Goal reconstruction The authors argue that humans can be interrupted from a task for indefinite periods of time during which the short term memory of the task has certainly been replaced, but after which the person quickly reconstructs the goal-space in order to complete the task. Teton employs a feature called short cutting to partially implement this capability.
Situated action The authors argue that many tasks humans perform do not involve planning but actions based on the situation. One example is canoeing in which the helmsman examines the upcoming rapids just before arriving and steers the boat accordingly. Analysis afterward might yield the conclusion that the behavior of the helmsman was planned when it actually was not.
Planned action The authors not only argue that planned action is necessary, they claim that the focus of research on classes of tasks that require planning has obscured the importance of situated action in humans.
Non LIFO goal selection The authors argue that a LIFO goal stack is unhuman using two examples. First the example of the "evening chores" states that if an individual's goal stack were LIFO, then after eating breakfast, the goal "Clean breakfast dishes" would be pushed on the stack. Upon arrival home, the goal "Prepare dinner" would be pushed on the stack. LIFO requires that dinner be prepared first and this is commonly violated.
In the second example, the authors performed a rational analysis of the columnar subtraction task and determined all possible LIFO-successful goal structures. Several elementary students were given subtraction problems which were to be solved using the columnar method. Eight of the students performed the subtractions using techniques that could not be solved by a LIFO goal structure and which were consistently employed by the students.

Common boundary properties for cognitive architectures in AI, the authors argue, are innateness, subject-universality and cognitive impenetrability. They argue that goal reconstruction is rightly an architectural capability that is subject-universal but neither innate nor cognitively impenetrable. They generalize the arguement to something like the following.

Innateness

In apposition to the notion of innateness, the authors argue that most cognitive theorists simply require that capabilities be acquired prior to attempting the task. Requiring that a capability be innate overly and inappropriately constrains the architect.

Universality

Since many architectural capabilities can be altered through learning, few are cognitively impenetrable and because the learning experience is unique, universality is abstract at best.

Impenetrability

This property of an architectural capability, the authors argue, can (and probably will) force uneconomical high-level processes. Using the example of goal reconstruction, they argue that this process has most of the features of the architecture's interpreter which is impenetrable. Therefore one is left with either the preclusion of goal reconstruction, or the unparsimonious and self-contradictory solution of producing an augmented form of the interpreter.

Return to the Top of TETON. Return to the Table of Contents.