anytime algorithms for learning anytime classifiers saher ... - Technion
anytime algorithms for learning anytime classifiers saher ... - Technion
anytime algorithms for learning anytime classifiers saher ... - Technion
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>Technion</strong> - Computer Science Department - Ph.D. Thesis PHD-2008-12 - 2008<br />
Test costs<br />
a1<br />
a2<br />
a<br />
a4<br />
a5<br />
a6<br />
40<br />
10<br />
10<br />
10<br />
10<br />
10<br />
T1 T2<br />
a1<br />
40 60<br />
- +<br />
a1<br />
20 80<br />
a2 a3<br />
15 5 10 70<br />
- + - +<br />
Figure 5.9: An example of applying cost discounts when <strong>for</strong>ming a repertoire <strong>for</strong><br />
interruptible classification. The numbers represent the number of examples that follow<br />
each edge. Because the value of a1 is required at the root of T1, any subsequent tree<br />
can obtain this value at no cost. The probability <strong>for</strong> testing a2 is 50% in T2. When<br />
inducing T3, the attribute a2 has already been measured with a probability of 50%.<br />
Hence, we discount the cost of a2 by 50% ($5 instead of $10). Similarly, the cost of<br />
a3 is discounted by 80% ($2 instead of $10).<br />
previous trees.<br />
Consider, <strong>for</strong> example, the trees in Figure 5.9. The probability to measure a1<br />
in T1 is 100%. There<strong>for</strong>e, when building subsequent trees, the cost of a1 would<br />
be zero. The probability <strong>for</strong> testing a2 is in T2 20%. Hence, when inducing T3,<br />
we discount the cost of a2 by 20% ($8 instead of $10). Similarly, the cost of a3 is<br />
discounted by 80% ($2 instead of $10).<br />
Because the trees may be strongly correlated, we cannot simply calculate this<br />
probability independently <strong>for</strong> each tree. For example, if T3 in the a<strong>for</strong>ementioned<br />
example tests a2 <strong>for</strong> 70% of the examples, we would like to know <strong>for</strong> how many<br />
of these examples a2 has been tested also in T2. There<strong>for</strong>e, we traverse the<br />
previous trees with each of the training examples and mark the attributes that<br />
are tested at least once. For efficiency, the matrix that represents which tests were<br />
administered <strong>for</strong> which case is built incrementally and updated after building each<br />
new tree.<br />
We refer to this method as discount repertoire. The repertoire is <strong>for</strong>med using<br />
the same method in Figure 5.5 with a single change: be<strong>for</strong>e building each tree,<br />
cost discounts are applied; the discounts are based on the trees already in the<br />
repertoire. Figure 5.10 <strong>for</strong>malizes the procedure <strong>for</strong> updating test costs. During<br />
classification we iterate over the trees until interrupted, as described in Figure<br />
5.8.<br />
5.4 Empirical Evaluation<br />
A variety of experiments were conducted to test the per<strong>for</strong>mance and behavior of<br />
TATA in 3 different setups: pre-contract, contract, and interruptible. In Chapter<br />
110