Coevolution Learning: Synergistic Evolution of Learning Agents and Problem Representations
Lawrence Hunter, Rockville Pike Bethesda · 1996
This paper describes exploratory work inspired by a recent mathematical model of genetic and cultural coevolution. In this work, a simulator implements two independent evolutionary competitions which act simultaneously on a diverse population of learning agents: one competition searches the space of free parameters of the learning agents, and the other searches the space of input representations used to characterize the training data. The simulations interact with each other indirectly, both effecting the fitness (and hence reproductive success) of agents in the population. This framework simultaneously addresses several open problems in machine learning: selection of representation, integration of multiple heterogeneous learning methods into a single system, and the automated selection of learning bias appropriate for a particular problem. Introduction One clear lesson of machine learning research is that problem representation is crucial to the success of all inference methods (see, ...