Hometown: Mumbai, Maharashtra, India
Graduation date: Spring 2021
MORE | Summer 2020
Investigating the Impact of Environmental Features on Multi-Agent Communication
Traditional reinforcement learning methods have been challenging to apply to multi-agent settings, thanks to the nonstationarity of the environment, until recently. This project aims to employ such multi-agent RL methods to understand how communication is learned between multiple agents. Specifically, the project aims to investigate what features cause effective communication protocols to be learned by the agents, and why. A centralized actor-critic method was chosen to facilitate multi-agent learning. So far, it has been observed that making certain features of the environment observable to certain agents causes the agents to learn to communicate effectively, when they couldn’t before.
Mentor: Yu (Tony) Zhang