[ACT-R-users] Production Compilation Module in ACT-R 6

[Niels Taatgen]

On 27 Mar 2007, at 14:16, Marcus Heinath wrote:

> Hello,
>
> I’m looking for some more detailed information about the production
> compilation module in ACT-R 6, particularly with regard to the  
> following
> questions:
>
> (I) What are the dependencies between those parameters?
>
> - initial experience

Is set to 10 by default. Larger values slow down how quickly real  
experiences can overcome the prior values of utility set by the  
parent rules. In my experience you never have to change this value,  
because it has very little impact on the model outcome.


> - initial-efforts
> - initial-success
> - initial-failures

By default, new productions start with a probability of success of 0,  
and a cost of G. If you want something else, you can change these  
parameters to accomplish this. I never saw any reason to do this though.


> - production learning rate

This is the main important parameter of production compilation. It  
controls the speed with which the utility of new rules approach the  
utility of the parent rules. Values must be between 0 and 1, but only  
values in the 0.02-0.2 range make sense (although you can use a value  
of 1 for debugging purposes). I almost always end up with a rate  
between 0.05 and 0.1

> - threshold time

Also not a very crucial parameter. It makes sure that if there is a  
lot of time between two production rule firings, no rule will be  
compiled. There must have been some model for which this was  
relevant, and that prompted its inclusion. You might want to set it  
to a high value, because at times I have been stumped at times by a  
model that wouldn't compile because the two rules were above threshold.

Note that there is a new utility learning mechanism in ACT-R, and  
which means that there are slightly different parameters in play  
right now.



>
> (II) What are characteristic value ranges for each parameter (I  
> only know
> the default values) and what are their meaning/influence on the  
> simulation?
>
> For example, the initial experience has a default value of 10, so  
> does this
> mean that an novice user could be represented by a value of 1 and  
> an expert
> user by a value of 20?
>
> (III) Is it possible to define a starting point for the production
> compilation, e.g. after a 3x successfully fired sequences of  
> Production 01
> and Production 02, they are composed to Production 03?
>
> P01 --> P02
> ...
> P01 --> P02
> ...
> P01 --> P02
>
> Production Compilation starts
> P03 (P01, P02)
>

Utility learning is supposed to accomplish this. Although the rule is  
created at the first opportunity, it will only be used after it has  
been compiled a number of times.

>

===================================================
Niels Taatgen - Carnegie Mellon University, Psychology, BH 345B
Also (but not now): University of Groningen, Artificial Intelligence
web: http://www.ai.rug.nl/~niels     email: taatgen at cmu.edu
Telephone: +1 412-268-2815
===================================================