anytime algorithms for learning anytime classifiers saher ... - Technion

Technion - Computer Science Department - Ph.D. Thesis PHD-2008-12 - 2008
Average Size
4000
3500
3000
2500
2000
1500
1000
0 100 200 300 400 500 600
Time [sec]
IIDT(1)
IIDT(0.1)
Figure 3.31: Time steps for a single run of IIDT on 10-XOR
we randomly partition each training set into 25% initial set and 75% incremental
set. Note that unlike the other methods, the non-incremental IIDT gets only the
initial set of examples.
Figures 3.32 shows the results for the XOR-10 datsets. Both versions of
I-IIDT exhibit good anytime behavior. They exploit the additional time and
examples. Their advantage over the greedy algorithms ID3 and C4.5, and over
the batch anytime algorithm is clear. These results demonstrate the need for
combining anytime and incremental learning: the anytime incremental learner
succeeds where the greedy incremental inducers and the anytime batch learner
fail. Comparing the performance of I-IIDT-TS to that of I-IIDT-EE shows that
IIDT-TS is better for this non-noisy setup. When we repeated the same experiment
with artificially added noise I-IIDT-EE was superior, due to its ability to
avoid overfitting.
3.7.6 Comparison with Modern Decision Tree Learners
During the last decade, several modern decision tree learners were introduced.
Although these learners were not presented and studied as anytime algorithms,
some of them can be viewed as such. In what follows we compare our proposed
anytime framework to three such algorithms: bagging, skewing and GATree.
We first give a brief overview of the studied methods and then compare their
performance to that of our anytime approach.
63