Multi-task learning and transfer: The effect of algorithm representation

Abstract

Exploring multiple classes of learning algorithms for those algorithms which perform best in multiple tasks is a complex problem of multiple-criteria optimisation. We use a genetic algorithm to locate sets of models which are not outperformed on all of the tasks. The genetic algorithm develops a population of multiple types of learning algorithms, with competition between individuals of different types. We find that inherent differences in the convergence time and performance levels of the different algorithms leads to misleading population effects. We explore the role that the algorithm representation and initial population has on task performance. Our findings suggest that separating the representation of different algorithms is beneficial in enhancing performance. Also, initial seeding is required to avoid premature convergence to non-optimal classes of algorithms.