A Memetic Multi-Agent Demonstration Learning Approach with Behavior Prediction

Yaqing Hou, Yifeng Zeng, Yew-Soon Ong

Abstract: Memetic Multi-Agent System (MeMAS) emerges as an enhanced version of multi-agent systems with the implementation of meme-inspired agents. Previous research of MeMAS has developed a computational framework in which a series of memetic operations have been designed for implementing multiple interacting agents. This paper further endeavors to address the specific challenges that arise in more complex multi-agent settings where agents share a common setting with other agents who have different and even competitive objectives. Particularly, we propose a memetic multi-agent demonstration learning approach (MeMAS-P) with improvement over existing work to allow agents to improve their performance by building candidate models and accordingly predicting behaviors of their opponents. Experiments based on an adapted minefield navigation task have shown that MeMAS-P could provide agents with ability to acquire increasing level of learning capability and reduce the candidate model space by sharing meme-inspired demonstrations with respect to their representative knowledge and unique candidate models.