Learning in Games II
summary
Analyzes learning in extensive form games where players observe only terminal nodes and shows repeated play can converge to outcomes that are self‑confirming equilibrium rather than Nash because off‑path beliefs never get data. Builds an overlapping populations model with Bayes updating from non‑doctrinaire priors, discounted lifetime optimization and steady states, uses law of large numbers and dynamic programming to prove existence and characterize limits. Key claims: 1/T experimentation suffices to rule out non‑Nash limits, patient players converge to Nash or to a subgame‑confirmed refinement depending on game structure and information.
AI-written summary · about
chapters (14)
- 0:00Extensive Form Games
- 2:48Terminal Node
- 5:57Learning Outcomes
- 19:04unitary selfconfirm equilibrium
- 21:18selfconfirm equilibrium
- 29:44path of s
- 32:48coons theorem
- 35:11learning dynamics
- 37:46aggregate model
- 40:00steady states
- 45:16any limit
- 55:19example
- 56:42empirics
- 57:28open questions
related talks
12mLearning In The Presence Of Strategic Agents: Dynamics, Equilibria, And Convergence
55mLearning Minimax Estimators Via Online Learning
12mTheoretical understanding of learning through the computational lens
49mFast Rates for No-Regret Learning in General Games
20mThe Complexity of Infinite-Horizon General-Sum Stochastic Games
15mMulti-Agent Reinforcement Learning In Stochastic Games: From Alphago To Robust Control
47mRebel: Combining Deep Reinforcement Learning and Search for Imperfect-Information Games
06mDecoding the Maze: New Frontiers in Achieving Nash Equilibrium in AI Architectures