Publication Date:
2015-06-23
Description:
Learning white box models is a very challenging task that, if successful, it extracts knowledge from a process; this learning process allows developing embedded applications with the outcome of hybrid artificial intelligent models. In previous studies, both with crisp or low-quality data, it has been shown that learning white box models using Genetic Programming (GP) and Genetic Algorithm Programming (GAP) is still penalized with several well-known problems: bloat, over fitting and population diversity. This research describes two very simple and intuitive techniques to deal with the two former problems making use of heuristics. The underlying idea is to avoid or to limit the surplus computation needed to reduce the problems' effects with simple human like rules. In the case of bloat, a heuristic to deal with incoherent node sequences is proposed; in the case of over fitting, the models are allowed a small error and those performing with higher bias are then penalized. These simple techniques are evaluated in a carefully designed test bed, which enables the analysis of their behaviour. Results show the proposed bloat preventing heuristic, which enhances the results both in the genotype and in the phenotype landscapes, while the over fitting technique slightly improves the evolutionary process.
Print ISSN:
1367-0751
Electronic ISSN:
1368-9894
Topics:
Mathematics
Permalink