Download BinClass: A Software Package for Classifying Binary Vectors User's

Transcript
try more candidates until SCk < SCk+1
if SCk < SCk ,
let SCk = SCk and restart
Step 3. Repeat Steps 1 and 2 until the conditions are not met
To ensure that the program stops, a safety value is needed, which is
dened by the switch-F. Set the initial number of candidate classications
with the switch -a. The automatic search usually yields better results and
avoids bad local minima, but on the other hand the running times can be
very long. Finding the correct parameters is not very easy and requires some
test runs to see how the program behaves on a particular dataset.
Some of the calculations in the program involves logarithms which are not
valid when the centroid value is 1 or 0. Thus, the value of the the centroid
needs to be xed with a small constant . The value of can be set with the
switch -E. Normally the default value works well. Too small value of can
also cause numerical problems.
Instead of repeating GLA, we can use smarter techniques. One of these
is local search.
Local Search SC minimizer
There are basically two ways to apply LS in BinClass. The rst one is to
utilize only one search operator by using the command line switches -r1 to
-r6. Another alternative is to try all operators using command line switches
-r7 or -r8. The iteration count of the LS algorithm is controlled by switch
-j. Usually the number of iterations needed ranges from 100 to 5000. The
switch -r8 chooses the adaptive version of the LS algorithm.
command: classify
switches
-q quiet mode (no screen output)
-EFF set epsilon to FF
-nXX use non-automatic search with XX test candidates
-aXX use automatic SC minimizer with at least XX test candidates
-Llename load initial centroids from the le
-l save (log) centroids
-C choose partition by error metric instead of SC
33