next up previous
Next: 3.7 Neural Networks versus Up: 3. Results Previous: 3.5 Correlation Among Methods

3.6 Bagging versus Simple network ensembles

Figure 6 shows the Bagging and Simple network ensemble results from Table 2. These results indicate that often a Simple Ensemble approach will produce results that are as accurate as Bagging (correlation results from Table 3 also support this statement). This suggests that any mechanism which causes a learning method to produce some randomness in the formation of its classifiers can be used to form accurate ensembles, and indeed, Ali and Pazzani [1996] have demonstrated similar results for randomized decision trees.


  
Figure 6: Reduction in error for Bagging and Simple neural network ensembles as a percentage of the original error rate. Also shown (white portion of each bar) is one standard deviation for these results.
\begin{figure}\centerline{\psfig{figure=nn-random-vs-bagging-new.ps,height=7.0in,bbllx=12pt,bblly=36pt,bburx=560pt,bbury=778pt} }
\end{figure}



David Opitz
1999-08-24