[ACT-R-users] Data-based parameter estimation and model selection in ACT-R

[Rhiannon L Weaver]

Hi everyone,

I am a stats grad student at CMU and am working on a method for 
likelihood-based estimation of ACT-R sub-symbolic parameters, and model 
selection between competing models based on traditional statistical 
techniques like AIC, BIC and Bayes Factors, as well as measures of 
complexity for these kinds of algorithmic models. I am currently 
developing some C++ code that will be able to apply Markov Chain Monte 
Carlo (MCMC) methods to an ACT-R model with observed data and come up with  
posterior distributions of parameters and likely paths of action (ie  
production firing/chunk recall) for sub-populations or individuals based 
on data observed for these populations/individuals.  Right now, my code is 
in the early stages:

1) I am still working with terminal or non-learning models
2) The code is restricted to models where either:
	a) all observable data for an individual is collected at the very 
end of a model run or,
	b) The observable data for each individual specifies the exact 
path of that individual (ie, progression of production firing and chunk 
recall) through the model.
3) The code cannot (yet) handle ACT-R perceptual-motor modules.
4) The code doesn't handle production compilation (ie production 
learning) or chunk merging.  Basically, the code needs to know a priori 
all possible productions that could fire and all possible chunks that can be 
retrieved through any model run.

I am looking for some (relatively) simple models that I could use to test 
the code out and possibly try to help answer some interesting research 
questions using these likelihood-based methods.  If anyone knows of any 
models out there, or if anyone has a model they think would be appropriate 
and is willing to let me take a look at it, or has any questions, 
comments, or suggestions, please email me (address below).  

Thank you for your time,
Rhiannon Weaver
rlweaver at stat.cmu.edu
PhD student, Statistics
Carnegie Mellon University

--------------------------
10 reasons to Watch Farscape: 

http://www.teevee.org/archive/2004/10/15/index.html