88 B. Hodjat and H. Shahrzad
Fig. 10 Distribution of originating segments for 100 fittest candidates in top-layer of runs with
64 segments. Note that 56 of the 64 segments have no candidates represented in the top 100
In Fig. 7 , for example, segment 1 has a disproportionately large contribution to
the top layer. It is as if candidates that were trained on segment 1 had a better
potential to learn and generalize. By the same token, we believe that the resulting
candidates originating from different segments have a higher diversity. This needs
to be investigated and qualified further and we hope to get to it in our future work.
The results show that, whennincreases, the number of segments contributing
to the top candidates drops. For example, forn D 16 three of the segments do
not contribute at all to the top 100 candidates from the run. This number goes up
to 56 fornD 64. This is mitigated by other segments that seem to have a better
representation of the data and so allow for even better generalization and overall
results. There is a point, however, after which the size of the segments, regardless of
the makeup, cannot sustain the run (e.g., see the 64 segment run results in Fig. 4 ).
5 Conclusions
We described the nPool model for cross-validation in a distributed evolutionary
system with incremental fitness evaluations. The real-world experimental results for
this approach are promising. Here are some of the benefits of this method: