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 />
ρ c<br />
= $30<br />
?<br />
$30<br />
a5<br />
a2<br />
ρ c<br />
= $60<br />
candidate<br />
splits<br />
a1<br />
a2<br />
a<br />
a4<br />
a5<br />
a6<br />
40<br />
0<br />
20<br />
25<br />
0<br />
10<br />
ρ c<br />
= $20<br />
a6<br />
$40<br />
a5<br />
a2<br />
ρ c<br />
= $20<br />
ρ c<br />
= $60<br />
Figure 5.1: Exemplifying the recursive invocation of TDIDT$. The initial ρ c was<br />
$60 (left-hand figure). Since two tests (a2 and a5) with a total cost of $30 have been<br />
used, their current cost is zero and the remaining budget <strong>for</strong> the TDIDT$ process<br />
is $30. a1, whose cost is $40, is filtered out. Assume that, among the remaining<br />
candidates, a6 is chosen. The subtrees of the new split (right-hand figure) are built<br />
recursively with a budget of $20.<br />
introduce the pre-contract component in the TATA framework, which adopts<br />
TDIDT$.<br />
5.1.1 Top-down Induction of Anycost Trees<br />
Any decision tree can serve as an anycost classifier by simply storing a default<br />
classification at each node and predicting according to the last explored node.<br />
Expensive tests, in this case, may block their subtrees and make them useless.<br />
In the pre-contract setup, ρ c , the maximal testing cost, is known in advance and<br />
thus the tree learner can avoid exceeding ρ c by considering only tests whose costs<br />
are lower. When a choice is made to split on a test θ, the cost bound <strong>for</strong> building<br />
the subtree under it is decreased by cost(θ). Figure 5.1 exemplifies the recursive<br />
invocation of TDIDT$. The initial ρ c was $60 (left-hand figure). Since two tests<br />
(a2 and a5) with a total cost of $30 have been used, their current cost is zero and<br />
the remaining budget <strong>for</strong> the TDIDT$ process is $30. a1, whose cost is $40, is<br />
filtered out. Assume that, among the remaining candidates, a6 is chosen. The<br />
subtrees of the new split (right-hand figure) are built recursively with a budget<br />
of $20. Figure 5.2 <strong>for</strong>malizes the modified TDIDT procedure, denoted TDIDT$.<br />
TDIDT$ is designed to search in the space of trees whose test-cost is at most<br />
101