Sequential Information Design: Markov Persuasion Process and Its Efficient Reinforcement Learning
summary
Introduces a Markov persuasion process where a sender observes per-step outcomes, sends recommended actions as signals, and seeks to maximize cumulative sender reward while receivers act myopically; proves a revelation principle reducing signals to recommended actions and formulates a persuasiveness constraint over posteriors. Builds OP4, an online algorithm that mixes optimistic LSVI‑UCB value estimates with a pessimistic confidence set over priors and enforces persuasiveness for all distributions in that set, and proves sublinear regret under a regularity condition (tabular case gives explicit bounds).
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:50Outline
- 1:22Motivation: Ride Hailing in Remote Areas
- 5:28How to Signal a Bayesian Driver?
- 13:29Bayesian Persuasion
- 20:53Markov Persuasion Process
- 28:58Online Setting for Sender
- 30:05Persuasiveness via Pessimism
- 33:53Exploration via Optimism
related talks
45mAGT and Machine Learning
11mThe Unreasonable Effectiveness of Prompt Learning
29mReinforcement Learning with Human Feedback
41mTheory of Modern AI: Learning Theoretic, Game Theoretic, and Algorithmic Perspectives
19mUnlocking the Full Potential of LLMs: Advanced Prompt Engineering Techniques
25mPrompt Based Recommendation Systems: A New Era of Personalized Experiences
32mEvaluating probabilistic decision making by LLMs: a decision-theoretic framework and application...
53mIterative preference learning methods for large language model post training