Equilibrium Learning in Unbounded Stochastic Games with Gaussian Transitions

Key points

• Learning equilibria from data is difficult when several decision makers interact indefinitely and the system state has no prescribed bounds.1
• The proposed procedure uses a single uninterrupted trajectory to update Gaussian-process models of the drift and to solve successively refined finite games.1
• We prove a perturbation result showing that stability of equilibrium selection, together with uniform KL accuracy of Gaussian transition models, is sufficient for regional uniform convergence of the induced values and approximate Nash behavior in the original game.1
• For GP estimators, simultaneous confidence bounds yield a rate controlled by posterior uncertainty, kernel information gain, and joint-state dimension.1

Sources (1)

• [1]Equilibrium Learning in Unbounded Stochastic Games with Gaussian Transitions
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 04:15 AM

  We examine discounted general-sum stochastic games with compact action sets and an unknown nonlinear drift observed through Gaussian transition noise.

  Learning equilibria from data is difficult when several decision makers interact indefinitely and the system state has no prescribed bounds.

Extractive summary: sentences quoted from the sources.