Independent Reading · 2024

Uncoupled Learning of Differential Stackelberg Equilibria with Commitments

Robert Loftin, Mustafa Mert Çelikok, Herke van Hoof, Samuel Kaski, and Frans A. Oliehoek · AAMAS

Stackelberg GamesMulti-Agent LearningDifferentiable Games

Full text

Read the Paper

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

Paper Information

Authors
Robert Loftin, Mustafa Mert Çelikok, Herke van Hoof, Samuel Kaski, and Frans A. Oliehoek
Venue
AAMAS · 2024
Open official paper

Summary

This paper develops decentralized learning dynamics for differential Stackelberg equilibria when agents cannot access one another’s payoff functions.

Core Idea

Zeroth-order estimates from observed behavior can replace cross-agent payoff information in hierarchical learning dynamics.

Method

The authors analyze uncoupled gradient estimators, prove convergence under established conditions, and introduce an online mechanism for negotiating leader-follower roles.

Key Takeaways

  • Stackelberg learning can be decentralized without sharing private utilities.
  • Commitment supports role formation in cooperative but asymmetric interactions.
  • The convergence guarantees match earlier coupled methods under comparable assumptions.

Questions

  • How does observation noise affect zeroth-order equilibrium learning?
  • Can more than two agents negotiate stable hierarchical roles online?