From the 1 of 3 linked papers with an AI index.
3 papers
RENEW: Towards Learning World Models and Repairing Model Exploitation from Preferences
Logan Mondal Bhamidipaty, Mykel Kochenderfer, Subramanian Ramamoorthy
The paper introduces RENEW, a method that uses human preferences over imagined rollouts to correct model exploitation in offline model-based reinforcement learning, focusing fine‑t…
Imperfect World Models are Exploitable
Logan Mondal Bhamidipaty, Esmeralda S. Whitammer, David Abel +2
We propose a novel definition of model exploitation in reinforcement learning. Informally, a world model is exploitable if it implies that one policy should be strictly preferred o…
Repairing Reward Functions with Feedback to Mitigate Reward Hacking
Stephane Hatgis-Kessell, Logan Mondal Bhamidipaty, Emma Brunskill
Human-designed reward functions for reinforcement learning (RL) agents are frequently misaligned with the humans' true, unobservable objectives, and thus act only as proxies. Optim…