Reproduce “Best-Response Bayesian Reinforcement Learning with Bayes-adaptive POMDPs for Centaurs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mustafa Mert Çelikok, Frans A. Oliehoek, Samuel Kaski
The research introduces the Human-Machine Centaur (HuMaCe) framework, which models human-AI collaboration as a sequential game where an AI assists a bounded-rational human. By employing Bayes-adaptive POMDPs, the AI learns the human's subjective model and cognitive biases, enabling it to strategically "nudge" human decisions towards objectively better outcomes while preserving human agency, and identifies the belief alignment problem in partially observable settings.
Key insights
- A formal decision-theoretic framework (HuMaCe) can effectively model intricate human-AI interactions, including human overrides and subjective task models.
- AI can successfully infer unknown human cognitive biases and costs from interactive behavior, enabling strategic "nudging" to improve objective outcomes without requiring full prior knowledge of human parameters.
- The "belief alignment problem" arises as a critical challenge for centaurs in partially observable environments, where divergent subjective observation models can prevent shared understanding and hinder effective collaboration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.