Reinforcement learning team, in collaboration with Gaming and AI team
Advised by Ida Momennejad in collaboration with Katja Hofmann
May - August 2023
New York / Redmond, USA
- Studied human-AI alignment through analysis of agent behavior using unsupervised manifold learning (UMAP) on a large-scale human gameplay dataset ~100K games (and AI datasets) in a multiplayer game, Bleeding Edge.
- Built GPT (generative pre-trained transformer) based game playing AI agents and trained them for targeted behavior replication (using imitation learning) through player-style identification from the above analysis.
- Manuscripts in preparation for conference and journal submissions.
Link to the company website: www.microsoft.com/en-us/research/about-microsoft-research/