Independent Reading · 2026
Learning in Bayesian Stackelberg Games With Unknown Follower’s Types
Matteo Bollini, Francesco Bacchiocchi, Samuel Coutts, Matteo Castiglioni, and Alberto Marchesi · arXiv preprint
Stackelberg GamesOnline LearningIncomplete Information
Full text
Read the Paper
Use View PDF or Download PDF to read the full paper on this device.
Paper Information
- Authors
- Matteo Bollini, Francesco Bacchiocchi, Samuel Coutts, Matteo Castiglioni, and Alberto Marchesi
- Venue
- arXiv preprint · 2026
Summary
This paper studies online learning in Bayesian Stackelberg games when the leader initially knows neither follower payoffs nor the distribution over follower types.
Core Idea
The feedback available to the leader determines whether meaningful learning guarantees are possible when follower types are unknown.
Method
The paper proves a negative result under action-only feedback and develops a no-regret epoch-based algorithm when follower types are revealed.
Key Takeaways
- Action feedback alone can be insufficient for efficient no-regret learning.
- Type feedback makes it possible to learn relevant response regions over time.
- Unknown payoff structures create challenges beyond learning an unknown type distribution.
Questions
- What additional structure could make action-only feedback learnable?
- How robust is the method when types or payoffs evolve over time?