[ACT-R-users] computational basis of act-r

[Christian Lebiere]

There are several levels at which oscillation theory, and a dynamical 
systems perspective in general, can be relevant to ACT-R.  Perhaps the most 
common invocation of the idea concerns the binding together of units that 
are active at the same time.  As such, it suggests a neural mechanism for 
chunk creation.  At a somewhat higher level, it can be used to look at the 
dynamics of how brain areas activate each other over time.  From that 
perspective, the evolution of ACT-R from a system synchronized around the 
production cycle to one more asynchronously driven by a variety of events, 
internal and external, simply reflects the complexity of the possible 
dynamics.  At a longer time scale (the rational and social band), the 
evolution of knowledge over time can also be seen as a dynamical system.

As Casper Hulshof pointed out, a lot of dynamical systems work is at a 
level of computation far removed from ACT-R.  The lesson I took from ACT-RN 
is that constraints on neural implementation are quite valuable in guiding 
architectural developments, but taken too literally they will only drag you 
down to a myriad of implementation problems that have little to do with the 
general (external) behavior of the system.  There is a general lack of 
appreciation in some quarters for the concept of abstraction.  That is 
particularly surprising given that it is really a basic foundation of much 
of hard science.  Chemistry works quite well in terms of simple algebraic 
(symbolic) equations, and trying to do it with quantum physics tools would 
very quickly prove intractable.  Moreover, we learned from the theory of 
computation that you can't tell a book from its cover.  Very different 
computational paradigms have been shown to be fundamentally equivalent, but 
slightly different programs within the same paradigm can lead to 
fundamentally different results.  Thus, an ACT-R model might have similar 
dynamics to a particular kind of neural network, which itself might have 
very different dynamics from another kind of neural network.

As for the opinion that started this thread, it brought to mind 
Khrushchev's old quote: "History is on our side.  We will bury you." 
Prediction is a tricky business, as we all know.