Large-scale Multi-agent Decision-making Using Mean Field Game Theory and Reinforcement Learning
Loading...
Authors
ZHOU, ZEJIAN
Issue Date
2021
Type
Dissertation
Language
Keywords
Alternative Title
Abstract
The Multi-agent system (MAS) optimal control problem is a recently emerging research topic that benefits industries such as robotics, communication, and power systems. The traditional MAS control algorithms are developed by extending the single agent optimal controllers, requiring heavy information exchange. Moreover, the information exchanged within the MAS needs to be used to compute the optimal control resulting in the coupling between the computational complexity and the agent number. With the increasing need for large-scale MAS in practical applications, the existing MAS optimal control algorithms suffer from the ''curse of dimensionality" problem and limited communication resources. Therefore, a new type of MAS optimal control framework that features a decentralized and computational friendly decision process is desperately needed. To deal with the aforementioned problems, the mean field game theory is introduced to generate a decentralized optimal control framework named the Actor-critic-mass (ACM). Moreover, the ACM algorithm is improved by eliminating constraints such as homogeneous agents and cost functions. Finally, the ACM algorithm is utilized in two applications.