A Memetic Multi-Agent Demonstration Learning Approach with Behavior Prediction
Abstract
Memetic Multi-Agent System (MeMAS) emerges as an en- \nhanced version of multi-agent systems with the implementa- \ntion of meme-inspired agents. Previous research of MeMAS \nhas developed a computational framework in which a series \nof memetic operations have been designed for implementing \nmultiple interacting agents. This paper further endeavors \nto address the speci c challenges that arise in more com- \nplex multi-agent settings where agents share a common set- \nting with other agents who have di erent and even compet- \nitive objectives. Particularly, we propose a memetic multi- \nagent demonstration learning approach (MeMAS-P) with \nimprovement over existing work to allow agents to improve \ntheir performance by building candidate models and accord- \ningly predicting behaviors of their opponents. Experiments \nbased on an adapted mine eld navigation task have shown \nthat MeMAS-P could provide agents with ability to acquire \nincreasing level of learning capability and reduce the candi- \ndate model space by sharing meme-inspired demonstrations \nwith respect to their representative knowledge and unique \ncandidate models.
Authors: Yaqing Hou, Yifeng Zeng, Yew-Soon Ong
Published in: Teesside University Research Portal (Teesside University) (2016)