A research team led by Kim Song Ho, a section head at the Faculty of Automation Engineering, has developed a multi-agent reinforcement learning experiment system using a process simulator.
In recent years, reinforcement learning has been considered as a learning method for acquiring behavior rules that can be flexibly adapted to complex and diverse environments, and many multi-agent reinforcement learning methods have been developed to effectively perform, in cooperation with several agents, complex problems impossible with one agent.
The multi-agent reinforcement learning experiment system consists of a process simulator, a controller, and an IPC.
On the IPC's experimental program interface, a user can simulate a learning process in detail by randomly changing the multi-agent system configuration and reinforcement learning parameters.
The change of the convergence rate according to the mutual cooperation algorithm of the agents can also be observed.
The research team developed a new multi-agent collaboration algorithm and demonstrated its superiority over previously developed ones.
This system can be used to develop multi-agent collaborative learning algorithms and to simulate real processes.
It can also offer practical trainings on intelligent supervisory control to students majoring in management system engineering and generalization to educational units across the country.
© 2021 KumChaek University of Technology