Rational Analysis and Cognitive Architectures

This is a brief summary of the arguments in John R. Anderson's paper The Place of Cognitive Architectures in Rational Analysis and Herbert A. Simon's comments on Anderson's paper, Cognitive Architectures and Rational Analysis: Comment.

Anderson's main premise, which he calls the "Principle of Rationality", states that "the cognitive system optimizes the adaptation of the behavior of the organism". In other words, a cognitive system will optimize its behavior with respect to its goals. Following this premise, Anderson claims that the behavior of a system can be better explained by examining its environment than by examining the architecture itself.

The following six steps describe the "rational analysis" approach to describing the behavior of a cognitive system.

Simon, in turn, explains some difficulties with the rational analysis approach. Anderson assumes that an organism's goals are known, but, in practive, it is often difficult to ascertain these goals. Even if the goals are know, there is nothing about evolution that guarentees an organism will be globally optimized toward them. All evolution does is favor the organism which is more fit than its neighbor. This doesn't necessarily produce a globally optimal organism by any means. Finally, Simon argues that one must look at the architectural limitation to understand a behavior. Almost no real world task can be performed "optimally" because an organism will run up against it's architectural limitations in trying to do so.

Anderson gives three examples where he uses rational analysis to explain some signiture phenomena for an architecture better than the architecture itself does. Simon addresses these three examples as well.

First, Anderson looks at the "power law learning" phenomena which the SOAR architecture tries to explain. Data suggests that, for some tasks, performance is a power funtion of practice. SOAR explains this behavior through its "chunking" learning mechanism.

Anderson explains this through his rational analysis approach by first making several assumption. He assumes that the human memory system is attempting to function as an "information processing system" much like a library. He then assumes that the retreival of memories (or "books") is governed by current retrieval queues and a usage history of the memories. For each particular memory the probability that it is a desired memory is a function of these retrieval queues and its usage history. Memories are then retrieved in order of most probable first. Retrieval stops when either the desired memories have been retrieved, or when the cost of not retrieving a memory is less than the cost of continuing to search.

Anderson formulates an equation for determining the probability that a particular memory is a desired one. He then uses the "usage history" part of that equation to explain power law learning.

Simon claims, though, that most of the power of this example comes from the assumptions made, and not from the rational analysis approach itself. For example, Anderson derives the power law learning phenomena from the probability equations, but these are obtained through the use of architectural assumptions (how memories are retrieved by humans) instead of environmental assumptions.

Andersons second example is the "fan-out affect" which the ACT architecture tries to explain. Humands seem to have difficulty with retrieving specific facts about subjects when that subject has many separate facts already associated with it. Experiments show that as more facts are added about a subject the slower the retrieval of any one particular fact is. ACT explains this through its propositional networks, where each subject has a number of connections to other subjects to form concepts. In ACT, only one connection can be tested at a time, so the more connections there are the slower ACT is in finding the correct ones.

Anderson uses the same model of memory as used for the power law learning analysis above. This time, however, he focuses on the retrieval queues part of the equations. From these he derives the fan-out affect phenomena.

Simon points out the same problems with this example as he pointed out with the power law learning example. Again, most of the power comes from the assumptions, not rational anaylsis.

Andersons third example involves our tendancy to catagorize objects. PDP models achieve catagorization implicitly by manipulating the strength of connections between elements based on experience.

Anderson argues that we use this capability to help us predict unknown attributes of newly encountered objects. He claims that the world is catagorized independently from the fact that we use the catagories. What we are doing is trying to find out the correct catagories for specific objects. He shows us the "ideal" catagorization algorithm, which he cannot use because it is too computationally complex. Instead he uses an iterative algorithm, without telling us where he gets it, to obtain his results.

Simon points out that the algorithm that Simon uses is only one of many that could be used to achieve catagorization. It is not, in any way, an optimal algorithm. Simon also takes issue with Anderson's claim that nature has provided pre-catagorized objects in our environment, pointing out that there are many objects which don't fit nicely into one specific catagory.

Simon's comments on the rational analysis approach are not intended to dissuade us from examining the environment to help explain behavior. Rather, he wishes to prevent the examinatin of ONLY the environment. Simon claims that it is in examining how an architecture deals with running up against its limitations when trying to deal with the environment that we will learn about the workings of an organism. Both the environment and the architecture must be examined to understand an organism.

Click here to see a discussion of the dichotomy between architectural and rational analysis.
Back to List of Other Meta-level Discussions