What is it about?
We formalize the problem of "demonstration sufficiency": how AI agents can determine if they've learned enough from human demonstrations to perform a task well. We propose a method to tackle this based on probabilistic bounds on how much the agent expects its current policy might underperform an expert policy. We test our approach in simulated environments and with a user study, showing that this method allows for accurate agent performance, high training sample efficiency, and allowance for suboptimal training demonstrations.
Featured Image
Photo by Jeffery Ho on Unsplash
Why is it important?
Having an AI agent that can determine when its current training demonstrations are sufficient can lead to easier, safer, and faster human-AI collaboration. It takes some of the burden of ensuring successful agent training off the humans' shoulders and places it on the agent.
Read the Original
This page is a summary of: Autonomous Assessment of Demonstration Sufficiency via Bayesian Inverse Reinforcement Learning, March 2024, ACM (Association for Computing Machinery),
DOI: 10.1145/3610977.3634984.
You can read the full text:
Contributors
The following have contributed to this page
