Learning and Equilibrium Refinements
summary
Shows that when agents are long lived and derive experimentation from discounted dynamic programming, limits called patiently stable profiles must be Nash equilibria, but need not satisfy backward induction; with experimentation rates like 1/T versus 1/√T off‑path nodes are learned very differently. Uses overlapping‑generations Bayesian models, optimal stopping and the Gittins index to derive a type‑compatibility refinement in signaling games, and introduces terminal node partitions to explain why learning is not normal‑form invariant.
AI-written summary · about
chapters (7)
- 0:00Introduction to the topic
- 2:27Defining the learning model
- 7:56System setup and steady states
- 14:17Patient stability and Nash equilibrium
- 18:53Simple games and subgame confirmation
- 20:16Refinements and game dynamics
- 21:35Invariance and conclusion
related talks
45mAGT and Machine Learning
41mSpotlight Talks
47mRebel: Combining Deep Reinforcement Learning and Search for Imperfect-Information Games
44mAI Ecosystems: Structure, Strategy, and Regulation
56mOn the Curses of Future and History in Off-policy Evaluation in non-Markov Environments
38mRetrospective on multi-agent learning as a path to intelligence
42mRethinking AI Alignment and Evaluation under Heterogeneous Preferences
11mThe Unreasonable Effectiveness of Prompt Learning