Independent Reading · 2020

Online Bayesian Persuasion

Matteo Castiglioni, Andrea Celli, Alberto Marchesi, and Nicola Gatti · NeurIPS

Bayesian PersuasionOnline LearningNo-Regret Learning

Full text

Read the Paper

Use View PDF or Download PDF to read the full paper on this device.

Paper Information

Authors
Matteo Castiglioni, Andrea Celli, Alberto Marchesi, and Nicola Gatti
Venue
NeurIPS · 2020
Open official paper

Summary

This paper studies repeated persuasion against receivers whose unknown types may be selected adversarially over time.

Core Idea

The sender can compete with the best fixed signaling scheme in hindsight despite uncertainty about the receiver’s utility function.

Method

The authors establish a computational hardness result and design no-regret algorithms for full-information and partial-feedback settings.

Key Takeaways

  • Unknown receiver types turn persuasion into an online learning problem.
  • Sublinear regret is possible under both rich and limited feedback.
  • Statistical learnability does not automatically imply efficient per-round computation.

Questions

  • Which structural assumptions could make the no-regret algorithms computationally efficient?
  • How would strategic or boundedly rational receivers change the feedback model?