Citation Context

...ision-tree method that does not store a large number of training instances, and so as our base learner we use a novel variant of ID4 (Schlimmer & Fisher, 1986) (which is an online version of the ID3 (=-=Quinlan, 1986) -=-offline decision-tree algorithm). We present empirical evidence that our extensions to ID4 improve performance in single trees and are critical to good performance in tree ensembles—our results supp...

Citation Context

...ds for invoking a “base” learning algorithm multiple times and for combining the resulting hypotheses into an ensemble hypothesis. We explore online variants of the two most popular methods, bagging (=-=Breiman, 1996-=-a) and boosting (Schapire, 1990; Freund, 1995; Brieman, 1996b). To our knowledge, all previous empirical evaluations of ensemble methods have taken place in offline learning settings (Freund & Schapir...

Citation Context

...e, 1990; Freund, 1995; Brieman, 1996b). To our knowledge, all previous empirical evaluations of ensemble methods have taken place in offline learning settings (Freund & Schapire, 1996; Quinlan, 1996; =-=Bauer & Kohavi, 1999-=-; Dietterich, in press)—our evaluation demonstrates similar online performance gains and also shows that ensemble methods are useful in meeting tight resource constraints. 3.1 Online Approaches to Ens...

Citation Context

...at are known to contain predictive information include register bits and branch target address bits; however, current methods for utilizing this information are table-based, e.g., (Heil et al., 1999; =-=Nair, 1995-=-). Our intended contribution to branch prediction (and to the design of other architecture-enhancing predictors) is to open up the possibility of using much larger feature spaces in prediction. 3 Onli...

Citation Context

... microprocessor simulator provided by the SimpleScalar 2.0 suite (Burger & Austin, 1997). We focused our study on eight branches from three different benchmark programs in the SPEC95 benchmark suite (=-=Reilly, 1995)�-=-�� the branches were selected because previous experiments indicated that single tree learners were performing significantly worse than Bayes optimally on these branches. Table 1 provides information ...