.po 1i
.ll 6.5i
.ce 10
\fB\s16Homework 3: AQ\fP\s0
.sp
.sz 14
CS 395T: Machine Learning
.sp .5
Due: Thursday, March 9
.sp
.ce 0
.pp
A version of the AQ algorithm for learning from examples which uses the
VERSION-SPACE system to compute stars is in the file AQ-VS.  It can be tested
on the examples in FIGURE-DATA or WEATHER-DATA. However, it is too inefficient
to run even on the reduced soybean dataset since star generation is very
explosive.  A version of AQ which uses beam-search to compute bounded stars is
in the file AQ and this system can be tested on large data sets if the
beam width (*max-star*) is set low enough.  AQ has AQ-TEST and
AQ-CATEGORIES functions analogous to the ones in VERSION-SPACE and ID3.
.sh 1 "System Modifications"
.pp
Hand in a commented version of the code you write to make the following
changes to AQ.
.sh 2 "Fuzzy Matching"
.pp
You are to change the matching of learned descriptions to test events in order
to allow "fuzzy" matching.  Use the approach from the original soybean
experiments as discussed in class.  This results in a numerical value for how
well a rule matches an instance.  When testing a new instance, simply assign
it to the class whose match value is the highest. 
.sh 2 "Truncating Covers"
.pp
It has also been found that simplifying a cover by truncating certain of its
complexes can improve performance.  Write code to simplify learned covers by
eliminating all of their complexes except the one that covers the largest
number of training examples.
.sh 1 "Experiments"
.pp
Run the initial AQ system on your personal location concept and briefly
comment on the results.  How does the learned DNF formula differ from
the tree learned by ID3?  How does performance compare to ID3?
.pp
The changes above really only make sense for multi-category problems.  Convert
your location data into a 2 category problem (places I like, places I don't
like) and run it using AQ-CATEGORIES instead of AQ-TEST.  In this case AQ
learns two covers instead of one: one for places you like, one for places you
don't. For this version of your location data and for the soybean data,
run the following cases:
.(l
Standard AQ
AQ with fuzzy match
AQ with truncation
AQ with both fuzzy match and truncation
.)l
Turn in a table of your results and some commentary on the results.

