Personal evaluation of ACT-R models

One of the ways to engage students to think more about ACT-R is to provide a way for them to see how well the theory of the mind corresponds to their personal experience. My approach to this is to start with something really basic like reaction time. Gunzelmann et al (2005) developed a well explained ACT-R model of Psychomotor Vigilance Task (PVT) and worked out the parameters to fit the their model to data from 88 hours of total sleep deprivation. With Terry Stewart's efforts this the ACT-R model was translated into PythonACT-R and so is available for students.

I have extended the model to include age and non-verbal IQ based on a brief review of the literature where both of these variables seem to operate through the speed of the production processing cycle of about 50 miliseconds. Age after 20 seems to slow the cycle and non-verbal IQ above 100 is associated with quicker reaction times and by inference faster processing cycles.

So the plan is to have students adjust the PVTdemoAGEPIQ.py model to make it specific to their individual situation and then run some simulations to see what the model predicts for situations like there own. To evaluate the prediction what is needed is a way for them to do the PVT and generate some comparable data. It is possible to use the environment part of PythonACT-R to collect some data but that requires a working ccmsuite simulation environment. As a backup to that situation (which is not possible with the standard ACT-R 6.0 simulator - it will run the task but not collect data as I understand it) I have modified the online reaction time tester by Jim Allen to be pretty close to the standard laboratory PVT equipment. This can be use to demonstrate the PVT situation with only ten trials (redemo10.html) or for actually collecting some comparison data with the 100 trial version (rtdemo100.html). Both situations: the PythonACT-R and the web page version produce a mean and standard deviation for the reaction times so comparisons can be straight forward. The difficulty of using an unstandardized version of the PVT is that at this point it may need to be calibrated but that raises the issue of experimental errors in replication which is a good discussion topic for students.

The teaching plan is that this "compare yourself to the PythonACT-R model" will provide both some interest and some challenges to students as they try to understand why the results came out the way they did. One interesting issue that I do not know how to address in a model is the consistant finding the women have slower reaction times by about 40 milliseconds and make fewer false alarms. I am hopeful that this will replicate in the class results and that it will lead to an interesting discussion of just what women are doing differently in the PVT situation (researchers have speculated that the difference is due to men an women using different "cognitive strategies" on the same task with the same instuctions -- "as fast as possible").