Modeling Human Architectures

<TITLE>Metalevel Discussion of Architectures</TITLE>
<pre>
<i>Jump to the <a href="topics"><b>topics list</b></a> page.<p></i>
<i>Jump to the <a href="../index.html"><b>Home</b></a> page.<hr><p></pre></i>
<h1>Modeling Human Architectures</h1>

When the world of artificial intelligence was somewhat more naive, every new
architecture was a new view of how the mind worked.  If only it would run fast
enough, then we would have thought in a box.<p><p>
These days the goals
of cognitive architectures are somewhat less lofty.  Still,
a few systems are using their architectures to test a cognitive model
as much as to get some functionality.  Both Brooks' subsumption architecture
and Soar come to mind.  The subsumption architecture goal is to rebuild human
intelligence from the ground up, and Soar can be considered a true 
implementation of Newell's cognitive bands for general intellience. 
These models render the systems more fragile(it is more 
difficult to integrate a particular capability) but the consistency of such 
models is to be admired and emulated.   
<p>
<p>However, these two architectures are in the minority.
In general the systems we have considered do not consider 
themselves a cognitive model.  That is, there is no unifying structure of 
the reasoning processes of the model which subsumes all of the working 
elements in an agent and its behaviors.  The individual architecture 
elements are assembled to implement some particular capability or to 
perform some particular task, and then the model is made.<p><p>
In particular, 
few systems we have considered can truly make analogies between their 
reasoning systems and human cognitive systems.  Psychological validity 
of a system, or a model which was created for that system, seems to be 
an afterthought in most cases.  It is often the case that an author of a 
particular architecture will call attention to a few unique examples of 
behavior in the system which on the surface emulate human reasoning, but 
the domains of these kind of behaviors typically turn out to be very 
limited, and overall not significant.  <p><p>
In conclusion, I would stress that human cognition needs to play a greater
role in the design of systems meant to embody general intelligence.  
Even though such an integrated model is difficult to implement, surely the
benefits of such a system would become clear in the long run, because it 
seems more likely that general intelligence could be implemented with a
tightly integrated system rather than a set of functional elements which do 
just that, 
function.