alue,

ߴ^∗ൌargmin

௠

ሼߴ௠, ∀݉∈ሾ1, ܯሿሽ

(8.7)

classification problem, ߝ௠ can be replaced by a classification

measurement such as the area under an ROC curve (AUC).

it is denoted by ߱௠. The fitness measure is revised as below,

ሺℓ௠ሻ is the sigmoid function of the size of the m^th RPN

ome,

߮௠ൌߙൈ߱௠൅ሺ1 െߙሻൈሼ1 െߪሺℓ௠ሻሽ

(8.8)

above equation, both 0 ൑߱௠൑1 and 0 ൑1 െߪሺℓ௠ሻ൑1 are

ed. The above fitness must be maximised to discover the best

scribed by an RPN chromosome, i.e.,

߮^∗ൌargmax

௠

ሼ߮௠, ∀݉∈ሾ1, ܯሿሽ

(8.9)

the fitness measurements of M models expressed by M RPN

omes have been measured, the M RPN chromosomes are sorted

ending order for a regression problem and in a descending order

sification problem.

M RPN chromosomes are divided into two parts. One part with the

fitness measurements is named as an elite set of RPN

omes and the corresponding models are called the elite models.

of RPN chromosomes are called non-elite ones and their

nding models are called non-elite models. The task which follows

the elite chromosomes to create or breed new chromosomes to

he non-elite chromosomes. Doing this is in the hope that newly

d chromosomes may produce new models with better fitness

ments than the old non-elite models. Therefore the next task is

breed new RPN chromosomes based on the elite RPN

omes.

ach iteration of a GP learning process, either the mutation

, or the dual-chromosome crossover operation or the single-