Speaker:
Dr. Sam Devlin
Microsoft Research Cambridge

Date:
15th July 2021

Title:
Coordinated Self-Play to Ad-Hoc Teamwork In Bleeding Edge

Abstract:
In collaboration with Ninja Theory, we are exploring multi-agent learning in their latest game, Bleeding Edge, which is a perfect testing ground for reinforcement learning agents trained to collaborate in teams. Whilst past work has often assumed we will be in control of all agents, if we want our agents to play well with any human they need to adapt quickly online. In this talk I will formalise the problem of ad-hoc teamwork and present our proposed approach to meta-learn policies robust to a given set of possible future collaborators. Then talk about recent work on imitation learning to provide human-like agents to train our ad-hoc teamwork enabled agent.