Joseph Suarez, Massachusetts Institute of Technology (MIT)
Talk Date and Time: February 21, 2023 at 4:00 pm - 4:45 pm EST followed by 10 minutes of Q&A on Zoom and IRB-5105
Topic: Making massively multiagent RL simple: Lessons from running competitions on Neural MMO
Abstract:
Neural MMO is a massively multiagent environment simulation platform for agent-based learning research. The project began 5+ years ago as a starting point for determining what types of innovations might be required for reinforcement learning methods to scale to more cognitively realistic environments. The platform now has several hundred users and has been the focus of the two largest reinforcement learning competitions in recent years. I will discuss the largest challenges overcome in the development of Neural MMO and how these lessons can help us transition RL research towards more interesting, complex environments by default.
Bio:
Joseph Suarez is a 4th year PhD student in Phillip Isola's lab at the Massachusetts Institute of Technology focused on scaling agent-based learning methods to complex, realistic environments. He graduated from Stanford with a degree in CS and AI, and interned at OpenAI for 6 months where he created the initial version of Neural MMO. Since then he has been developing Neural MMO and working with industry sponsors including AICrowd, Parametrix.ai, Stability.ai, and Carper AI to run competitions for Neural MMO at IJCAI 2022 and NeurIPS 2022. He also runs the monthly Learning in Foundation Environments (LIFE) talks which present talks from researchers and developers working on new tools, platforms, and infrastructure for reinforcement learning environments.