nPool: Cross-Validation in EC-Star 81
Fig. 2 Evolution Coordinators make copies of candidates and send them to the Evolution engines
for further aging. The reported results are merged. Note that, in step 8, if the candidate is not worthy
of being added to the new layer, it is discarded, but its ‘dead’ copy may remain in lower layers to
help with filtering (step 7)
such that within each iteration a different fold is held-out for validation while the
remainingk 1 folds are used for learning. The learned models are asked to make
predictions about the data in their respective (unseen) validation folds.
2 Description
Similar to k-fold cross validation, in nPool, the training data sample set is divided
(roughly) equally into nsegments. This division can be done randomly and
segments should be mutually exclusive. Each evolution engine validates candidates
exclusively on data samples from the segment randomly assigned to them by
the Evolution Coordinator at start up. Every candidate generated by an Evolution
Engine is tagged with the segment assigned to that Evolution Engine. We call this
the candidate’s originating segment, and at birth, the candidate’s current segment
is said to be its originating segment. Once a candidate has been evaluated on a
sufficient number of data samples from its own segment, it is said to have hit
its graduation age. The Evolution Coordinator tags graduated candidates with a
segment id (out of thenavailable) that they have not been tagged yet. This id is
the new current segment for the candidate.