SOAR is the most recent cognitive architecture that was proposed by Newell and his colleagues. The architecture is quite sophisticated and we will gloss over much of the detail that is required to deeply understand and evaluate this proposed architecture. The bibliography provided at the bottom of this page provides references to the more complete discussions of this architecture.

    Soar is a production rule system, but one that is quite differently constructed than the one that we have already studied. The table below taken lists six aspects of the Soar architecture that Newell identifies as the main characteristics of SOAR. The first characteristic is quite notable. Within a typical program there are almost always segments of the program that are simply sequences of instructions. In Soar there is no such code. Every task is a 'problem' which is solved via search in the appropriate problem space. This property is referred to as Universal Subgoaling. For Soar everything is a problem in some problem space.

   Note that the second characteristic states that all of the systems knowledge is stated in the form of production rules. Soar is thought of as a kind of massive memory whose content is accessed through the condition portion of the production rules. Consequently Soar utilizes a variant of the Recognition Act Cycle that was discussed earlier. However, in this case the authors of Soar refer to am Elaboration Decision Cycle. The "working memory" has become in Soar a much more elaborate temporary memory structure that holds information pertaining to the state of the current "problem solving context." The technical details will be ignored. But think of this context as including information relevant to the decision about what problem space to use for the problem, what states of the space might be under consideration, what aspects of an operator might be under consideration, and so on. In other words the entire process has been opened up and production rules relevant to any of these aspects of the problem might be retrieved during the Elaboration Phase. During the Elaboration phase all productions that are accessed from information in this current "problem solving context" are retrieved.

   The fourth characteristic mentioned below refers to a preference language. During the Decision Phase the rule that is to be applied must be identified. The preference language is used to express any preferences that apply to the set rules that have been instantiated during the Elaboration Phase. If these is no clear preference, then the system is at an impasse. There are several ways in which an impasse can be encountered. Regardless, of how the impasse is encountered, the system attempts to resolve the impasse. Characteristic 5 below refers to this aspect of the system. Essentially the system creates a subgoal which is to resolve the impasse. For example, if it doesn't know which operator to apply in a state space search, the subgoal of providing-a-basis-for-choosing-among-the-operators would be created. Successful solution to this impasse-generated subgoal would lead to a form of learning that the authors refer to as chunking. Note that learning is now an intrinsic part of the architecture; not a separate module or function.

 

 
   The figure below lists the properties of this learning which is referred to as Chunking. Note that the result of the learning is the creation of knowledge that is in the form of production rules. The production rule results from a process that constructs the rule from the information that is held in what I have referred to as the "problem solving context." Because of the way in which the production rule is formed it will become available and dominant in the problem solving context in which it was acquired (see 2 and 4 below); and possibly in somewhat different contexts as well (see 3 below).
 

 
     

Bibliography

Laird, J., Rosenbloom, P., & Newell, A. Universal Subgoaling and Chunking: The Automatic Generation and Learning of Goal Hierarchies. Boston: Kluwer, 1986.

Rosenbloom, P.S. & Newell, A. The Chunking of Goal Hierarchies: A Generalized Model of Practice. In R.S. Michalski, J.G. Carbonell, & T. Mitchell (Eds.), Machine Learning. Vol. II. Los Altos, CA: Kaufmann, 1986. Pp. 247-288.

Newell, A. Unified Theories of Cognition. Cambridge, MA: Harvard University Press, 1990.

Human Cognition - Table of Contents
 © Charles F. Schmidt