Session: DAC-09-1: Design of Complex Systems

Paper Number: 143326

143326 - Impact of Social Learning on Teamwork Efficiency in Learning-Based Self-Organized Systems 

Self-organized systems (SOSs) possess the great potential to adapt to dynamic environments, and thus, advance the level of autonomy in engineering systems. Learning-based approaches, particularly multiagent reinforcement learning (MARL), have empowered the design freedom of SOSs while enhancing the system to be more intelligent. This is primarily because MARL encourages agents to learn from past failures and data through training. However, the training process is conducted in a “black box”, whose embeddings are uninterpretable for humans yet. This poses significant challenges for system designers, especially when task complexity reaches a certain extent, and agents experience a shortfall in learning due to limited available information. Therefore, in order to address how to steer more effective learning, we broaden the views of individual agents, enabling them to recognize the presence and actions of other members within a certain range of distance. This approach fosters social learning within AI agent groups. Prior work, however, has shown the intricacy of the social mechanism. Inappropriate social abilities can undermine a team’s performance. Thus, it is imperative to delve into the effects of social learning to facilitate the system design, further narrowing down the understanding gap between humans and AI agents. In this study, we train agent teams to engage in social learning to complete an assembly task involving collision avoidance. We aim to answer how social abilities influence teamwork at both the team and individual agent levels in terms of energy utilization. By examining the impact of team energy efficiency and teamwork division, our findings indicate a mapping linking appropriate social abilities to tasks of a certain complexity level. Suitable social abilities contribute to elevated team intelligence, more effective teamwork division, and higher energy efficiency, while insufficient social abilities may hinder cooperation and reduce energy efficiency.

Presenting Author: Bingling Huang California State University, Fullerton

Presenting Author Biography: Dr. Bingling Huang is currently an Assistant Professor at California State University, Fullerton. She earned her Ph.D. in Mechanical Engineering from University of Southern California (USC) in 2023, and M.S. in Computer Science from the same university in 2021. Prior to that, she completed her M.S. in Aeronautical Engineering and B.Eng. in Aircraft Design and Engineering, both from Beihang University in 2018 and 2015, respectively. Dr. Huang’s research focuses on engineering design, complex and self-organizing systems, knowledge acquisition, and multi-agent reinforcement learning and machine learning for industry applications.

Authors:

Bingling Huang California State University, Fullerton
Hao Ji University of Southern California
Yan Jin University of Southern California