Presentation about networked MARL systems
Date:
The slides (.pptx) can be downloaded here. slide
Discussed with Assistant Prof. Zongqing Lu.
Abstract
As the basic framework for solving marl problems, centralized training and decentralized execution (CTDE) greatly prospers a series of algorithms, such as QMIX, MADDPG, and MAPPO. However, centralized training faces two fundamental problems, exponential increasing state-action space and overfitted strategy under the centralized information structure. Both are crucial issues for deploying the MARL algorithms from the simulator to practical applications. In this report, we will focus on the decentralized training and decentralized execution (DTDE) framework for networked MARL systems, including the definition of networked MDP, the brief survey of DTDE papers, our previous attempt, and some discussions.
Related Works
QD-Learning. TSP 2013. Kar.
ConseNet. ICML 18. Zhang Kaiqing.