Multi-agent Reinforcement Learning Project

For more information, you can check this Repository in my Github.

This is my personal research project, utilizes the PyMARL framework to setup environment Matrix Game and Predator Prey, allows multi-agent play games in these environments, to demonstrate the patholog. The project’s technological stack comprises of Python and PyTorch

I show the influence of the relative overgeneralization pathology for the VDN and QMIX algorithms, a phenomenon where an agent’s behavior becomes overly influenced by the actions of its teammates, leading to suboptimal decision-making. Then I modified these algorithm to address the pathology.

• VDN (Value-Decomposition Networks) is a multi-agent reinforcement learning algorithm that decomposes the centralized value function into a sum of individual value functions, enabling agents to make decentralized decisions.
• QMIX (Q-value Mixing) is a variant of VDN that uses a monotonic mixing function to improve the performance of the algorithm.

This figure showes relative overgeneralization pathology in multiagent learning. (Wei and Luke, 2016)

• This project used PyMARL framework to setup environment Matrix Game and Predator Prey, allows multi-agent play games in these environments, to demonstrate the patholog.

This is a structure diagram for better understanding PyMARL

• In the experiments, I demonstrated that the two algorithms, VDN and QMIX, were indeed affected by RO.

The RO pathology occured when agents play Predator Prey game

I then made modifications to the two algorithms within this project. Finally relieved the RO pathology.