nPool: Cross-Validation in EC-Star 85
4 Experiments
For the purposes of the experiments in this paper, we split the training data into 1,
2, 4, 8, 16, 32, and 64 mutually exclusive segments respectively. The data sets are
regenerated randomly from the training set before each experiment.
All experimental runs used the same number of Evolution Engine processing
nodes, namely 64, with a run duration of 24 h. The maturity age, and consequently,
the age-layer range, in all experiments was set to 150. The age-layer quota for all
runs was set at 100 candidates per age-layer. The Evolution Engine pool size for all
runs was fixed at 1000 candidates, with an elitist percentage of 10 %.
For the nPool experiments, the top-layer min age for the runs is reduced in inverse
proportion to the number of segments. For n=1, top-layer min-age is 30,000. For
n=2, it is set to 15,000, for n=4, it is 7500, for n=8, it is 3750, for n=16, it is 1800,
for n=32, it is 900, and for 64, it is 450. In all cases, of course, the harvest age is the
same (i.e., 30,000).
4.1 Results
In Fig. 4 we compare average fitness results from training. Note that the training
results include the results from running on the unseen segments in the cases where
nis greater than one. The unseen set is the same for all runs.
Fitness580,000600,000620,000640,000660,000680,000700,000720,000740,000760,000(^1248163264) n
Average Training Average Unseen
Fig. 4 Comparison of nPool training and unseen results between runs with 1 (i.e., no nPool), 2, 4,
8, 16, 32, and 64 segments for the top four candidates. These candidates are then run on the unseen
set withheld from training for comparison