Hybridizing novelty search for transfer learning

Abstract

This study investigates the impact of genotypic and behavioral diversity maintenance methods on controller evolution in multi-robot (RoboCup keep-away soccer) tasks. The focus is to examine the impact of these methods on the transfer learning of behaviors, first evolved in a source task before being transferred for further evolution in different but related target tasks. The goal is to ascertain an appropriate controller design (NE: Neuro- Evolution) method for facilitating improved effectiveness given policy transfer between source and target tasks. Effectiveness is defined as the average task performance of transferred behaviors. The study comparatively tests and evaluates the efficacy of coupling policy transfer with several NE variants. Results indicate a hybrid of behavioral diversity maintenance and objective-based search yields significantly improved effectiveness for evolved behaviors across increasingly complex target tasks. Results also highlight the efficacy of coupling policy transfer with the hybrid of behavioral diversity maintenance and objective based search in order to address bootstrapping and deception problems endemic to complex tasks.