Patricia Albacete's ISP Comps Questions List

Patricia L. Albacete

Questions for the Comprehensive Examination
April 11, 1995.


Question 1

Gordon Logan and others have observed that in some tasks, the power
law of practice only applies to individual problems and not to the
procedure or rule as a whole.  For instance, Logan gave subjects
practice answering questions of the form B+3=? where the answer is the
3rd letter in the alphabet after B (i.e,. E).   Thus, the correct answer
to N+5=? is S.  Subjects do this task by starting with the given letter,
and counting N further letters through the alphabet, where N is the
given digit (between 1 and 5).  With sufficient practice, their
performance becomes faster and faster, but only on the number-letter
pairs that they actually practiced!  Thus, if they practiced D+3=? then
that combination would get faster but E+3=? would not and D+4=? would
not.  The decline in reaction time fits a power law beautifully. 

Anderson and Fincham have observed that on certain tasks, the power law
of practice applies to the whole procedure or rule, and not just to
individual problems.   They gave subjects practice on simple procedural
rules of the form " is played on  at  and on 
at "  e.g., "Hockey is played on Sunday at 2 and on Monday at 4."
 For  a given sport,   was always a certain numbers of days before
or after , and  was always a certain number of hours before
or after .  For instance, for hockey,  was always one day
after  and  was always 2 hours after .  Anderson and
Fincham trained subjects by giving them fill-in-the-blank questions of
the form "Hockey is played on Tuesday at 10 and ___ at ___.  The subject
would fill the blanks by selecting "Wednesday" and "12" from menus as
fast as possible.  Anderson and Finham found that their reaction time
decreased according to a power law even though their training never
repeated combinations during training! That is, the stimulus triple
hockey-wednesday-12 would appear only once during training.  How can
ACT* explain both this result and Logan's?



Question 2

a) Suppose you wanted to build a tutoring system for chess.  You have
available an excellent chess program.  However, the chess program is
completely opaque to you.  In fact, it is just an address on the net --
you send it a chess position, and it returns the list of all legal
moves, each with a numerical ranking.  It runs on a computer size of the
 moon, so it does this in 200 microseconds. Anyway, you can't get at the
expertise that is embedded in this program.   Moreover, you do not have
decent natural language processing -- you're stuck with ordinary
command/menu input and template-driven output.

How would you go about building a tutoring system for chess?  What major
architectures (i.e., module-level designs) would you consider? How would
you obtain knowledge to encode in the tutor?  What kind of learning
would you target?  How would the tutor encourage that learning?  Would
you have the student do other activities besides just play chess?

b) Now suppose that you have in addition to the chess machine and
expert human chess tutor.   During the software development, you can
take videos of the tutor as she coaches students; you can ask her
questions; etc. but she will not be around when your ITS is trying to
tutor students, so you can't have her answer their questions, for
instance.  Now what would you do differently, and why?



Question 3

Describe Markov condition, its empirical foundations, and various
ways in which it can be used to reduce the amount of computation in 
probabilistic reasoning.



Question 4

Describe how you would apply probabilistic and decision-theoretic methods 
to student modeling in the context of intelligent tutoring systems.  Feel 
free to mention approaches taken by others in this domain but, most 
importantly, try to go beyond what has been done reflecting on your own 
knowledge of probability and decision theory. For each of the possible 
applications that you have identified, list potential problems and outline 
how these may be solved.



Question 5

A medical researcher believes that the ultraviolet part of sunlight kills
airborne germs. It follows that the amount of sunlight should be a major
determinant of the occurrence of airborne contagious diseases such as the
common flu. He writes a proposal to NSF to test this hypothesis with the 
help of weather data already collected by the US government as well as
medical information from the data banks of hospitals and clinics nationwide.
No new data are to be collected.

Design the data analysis for the study:

a) Under the assumption that the investigator has unlimited resources
(time, research assistants, computer power, etc.).

b) Under the assumption that the investigator has one research assistant 
and that the report has to be ready in six months.



Question 6

An educational researcher wants to investigate the effect of an ITS that
embodies an entirely new approach to teaching Algebra. S/he hypothesizes
that the system will be very beneficial for students who do not know
anything at all about Algebra. On the other hand, s/he also believes that 
it will confuse students who already know some Algebra, by forcing them to
restructure what they already know.

a) Design a correlational study to test the hypothesis. Explain the 
possible outcomes and their meanings.

b) Design an experiment study to test the hypothesis. Explain the possible
outcomes and their meanings.



Question 7

Freud once proposed that there are exactly two (major) determinants of
how happy and satisfied human beings are with their lives: Love and Work.
He meant that there are two basic inner needs --the need to be loved by 
other people and the need to be productive-- and that happiness is a function
of the extent to which those two needs are met. Design a study to test this
hypothesis.