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 />
Misclassification cost<br />
40<br />
35<br />
30<br />
25<br />
20<br />
15<br />
10<br />
5<br />
0<br />
0 1 2 3 4 5<br />
Sample size (r)<br />
C4.5<br />
TATA<br />
Figure 5.13: TATA with different sample sizes on the Multi-XOR dataset<br />
Besides being able to produce good anycost trees, TATA itself is an <strong>anytime</strong><br />
algorithm that can trade <strong>learning</strong> resources <strong>for</strong> producing better trees. Our next<br />
experiment examines the <strong>anytime</strong> behavior of TATA by invoking it with different<br />
values of r. Figure 5.13 shows the results. As we can see, all TATA versions are<br />
better than C4.5. With the increase in r, the advantage of TATA increases. The<br />
most significant improvement is from r = 0 to r = 1. TATA with r = 0 is a<br />
greedy algorithm that uses local heuristics. When r > 0, TATA uses sampling<br />
techniques to evaluate attributes and there<strong>for</strong>e becomes significantly better. For<br />
more difficult concepts, where only a combination of a large number of attributes<br />
yields in<strong>for</strong>mation, a larger value of r would be needed.<br />
5.4.2 Contract Classification<br />
TATA uses repertoires to operate in the contract and interruptible classification<br />
setups. To examine the anycost behavior of TATA in the contract setup we built<br />
3 repertoires, each of size 16. The first repertoire uses pre-contract-TATA(r = 0)<br />
to induce the trees, with uni<strong>for</strong>m contract gaps. The second repertoire uses precontractTATA(r<br />
= 3) to induce the trees, with uni<strong>for</strong>m cntract gaps. The trees<br />
in the third repertoire are also <strong>for</strong>med using pre-contract-TATA(r = 3), but with<br />
the hill-climbing approach instead of uni<strong>for</strong>m gaps.<br />
The repertoires were used to classify examples in the contract setup. 120<br />
uni<strong>for</strong>mly distributed ρ values in the range [0 − 120%ρ c max] were used as contract<br />
parameters. Figure 5.14 describes the per<strong>for</strong>mance of these three repertoires on<br />
4 datasets (Glass, AND-OR, MULTI-XOR, and KRK), averaged over the 100<br />
runs of 10 times 10-fold cross-validation. In addition, we report the results of the<br />
115