A method to increase the generalization ability of genetic programming (GP) is proposed in this paper. The idea consists in giving a second chance of mating to individuals belonging to “old” generations (hence the name of the method: “second chance GP”). Although original, the idea is inspired by well-known concepts such as short-term memory schemes, that have already been used in evolutionary computation so far. Three complex real-life applications characterized by a high dimen- sionality of the space of the features have been used to experimentally validate the approach. These three problems are interesting for our study because in one of them standard GP has no overfitting, in the second one standard GP tends to slightly overfit training data and in the third one standard GP seriously suffers from overfitting. The obtained results show that “second chance GP” performs in a comparable way to stan- dard GP when the latter does not overfit, while it clearly outperforms standard GP when the latter suffers from overfitting.
Reinsertion of Old Genetic Material: Second Chance GP
Luca Manzoni;
2011-01-01
Abstract
A method to increase the generalization ability of genetic programming (GP) is proposed in this paper. The idea consists in giving a second chance of mating to individuals belonging to “old” generations (hence the name of the method: “second chance GP”). Although original, the idea is inspired by well-known concepts such as short-term memory schemes, that have already been used in evolutionary computation so far. Three complex real-life applications characterized by a high dimen- sionality of the space of the features have been used to experimentally validate the approach. These three problems are interesting for our study because in one of them standard GP has no overfitting, in the second one standard GP tends to slightly overfit training data and in the third one standard GP seriously suffers from overfitting. The obtained results show that “second chance GP” performs in a comparable way to stan- dard GP when the latter does not overfit, while it clearly outperforms standard GP when the latter suffers from overfitting.Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.