Relational concept learning by cooperative evolution

Abstract

Concept learning is a computationally demanding task that involves searching large hypothesis spaces containing candidate descriptions. Stochastic search combined with parallel processing provide a promising approach to successfully deal with such computationally intensive tasks.Learning systems based on distributed genetic algorithms (GA) were able to find concept descriptions as accurate as the ones found by state-of-the-art learning systems based on alternative approaches. However, genetic algorithms' exploitation has the drawback of being computationally demanding.We show how a suitable architectural choice, named cooperative evolution, allows to solve complex applications in an acceptable user waiting time and with a reasonable computational cost by using GA-based learning systems because of the effective exploitation of distributed computation. A variety of experimental settings is analyzed and an explanation for the empirical observations is proposed.