... 0.¹

For those few cases where $\epsilon_k$ becomes 0 (less that 0.12% of our results) we simply use a large positive value, $log(\beta_k) = 3.0$, to weight these networks. For the more likely cases where $\epsilon_k$ is larger than 0.5 (approximately 5% of our results) we chose to weight the predictions by a very small positive value (0.001) rather than using a negative or 0 weight factor (this produced slightly better results than the alternate approaches in pilot studies).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... noise²

X% noise indicates that each feature of the training examples, both input and output features, had X% chance of being randomly perturbed to another feature value for that feature (for continuous features, the set of possible other values was chosen by examining all of the training examples).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.