anytime algorithms for learning anytime classifiers saher ... - Technion
anytime algorithms for learning anytime classifiers saher ... - Technion
anytime algorithms for learning anytime classifiers saher ... - Technion
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
<strong>Technion</strong> - Computer Science Department - Ph.D. Thesis PHD-2008-12 - 2008<br />
Average Size<br />
Average Accuracy<br />
40<br />
38<br />
36<br />
34<br />
32<br />
30<br />
28<br />
26<br />
24<br />
0 50 100 150 200 250 300<br />
72<br />
70<br />
68<br />
66<br />
Time [sec]<br />
ID3<br />
C4.5<br />
IIDT(1)<br />
IIDT(0.1)<br />
64<br />
ID3<br />
C4.5<br />
IIDT(1)<br />
62<br />
IIDT(0.1)<br />
0 50 100 150 200 250 300 350<br />
Time [sec]<br />
Figure 3.28: Anytime behavior of IIDT on the Glass dataset<br />
3.30 show the <strong>anytime</strong> per<strong>for</strong>mance of IIDT in terms of tree size and accuracy <strong>for</strong><br />
the Glass, XOR-10, and Tic-tac-toe datasets. Each graph represents an average<br />
of 100 runs (<strong>for</strong> the 10 ×10 cross-validation). Unlike the graphs given in the previous<br />
section, these are interruptible <strong>anytime</strong> graphs, i.e., <strong>for</strong> each point, the y<br />
coordinate reflects the per<strong>for</strong>mance if the algorithm was interrupted at the associated<br />
x coordinate. In the contract algorithm graphs, however, each point reflects<br />
the per<strong>for</strong>mance if the algorithm was initially allocated the time represented by<br />
the x coordinate.<br />
In all cases, the two <strong>anytime</strong> versions indeed exploit the additional resources<br />
and produce both smaller and more accurate trees. Since our algorithm replaces a<br />
subtree only if the new one is smaller, all size graphs decrease monotonically. The<br />
most interesting <strong>anytime</strong> behavior is <strong>for</strong> the difficult XOR-10 problem. There,<br />
the tree size decreases from 4000 leaves to almost the optimal size (1024), and the<br />
accuracy increases from 50% (which is the accuracy achieved by ID3 and C4.5) to<br />
almost 100%. The shape of the graphs is typical to those of <strong>anytime</strong> <strong>algorithms</strong><br />
60