Neuro-evolution behavior transfer for collective behavior tasks

Abstract

The design of effective, robust and autonomous controllers for multi-agent and multi-robot systems is a long-standing problem in the fields of computational intelligence and robotics. Whilst nature-inspired problem-solving techniques such as reinforcement learning (RL) and evolutionary algorithms (EA) are often used to adapt controllers for solving such tasks, the complexity of such tasks increases with the addition of more agents (or robots) in difficult environments. This is due to specific issues related to task complexity, such as the curse of dimensionality and bootstrapping problems. Despite an increasing attempt over the last decade to incorporate behavior (knowledge) transfer in machine learning so that relevant behavior acquired in previous learning experiences can be used to boost task performance in complex tasks, using evolutionary algorithms to facilitate behavior transfer (especially multi-agent behavior transfer) has received little attention. It remains unclear how behavior transfer addresses issues such as the bootstrapping problem in complex multi-agent tasks (for example, RoboCup soccer). This thesis seeks to investigate and establish the essential features constituting effective and efficient evolutionary search to augment behavior transfer for boosting the quality of evolved behaviors across increasingly complex tasks. Experimental results indicate a hybrid of objective-based search and behavioral diversity maintenance in evolutionary controller design coupled with behavior transfer yields evolved behaviors of significantly high quality across increasingly complex multi-agent tasks. The evolutionary controller design method thus addresses the bootstrapping task for the given range of multi-agent tasks, whilst comparative controller design methods yield scant performance results.