What Does Machine Learning Offer Game Theory (And Vice Versa)?
summary
Argues that machine learning can compress big strategic environments into tractable game abstractions and shows two working patterns: for social dilemmas a joint RL run produces a cooperative policy via joint RL and a separate minimax/punishment policy; partners are classified by forward-looking Q-value differences or retrospective payoff thresholds and the method reliably enforces cooperation in gridworld coin games and coordination tasks. For markets a valuation matrix gets low-rank factorization plus k‑means clustering into a low-rank market abstraction, solved by a convex program and lifted back; theory bounds error linearly in factorization and clustering loss and datasets show practical low-rank structure.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 2:00Poker AI
- 4:08Prisoners Dilemma
- 7:47State Space
- 12:58Coordination
- 13:44Market Design
- 14:25Equilibrium Allocations
- 21:53Theory
- 23:37Computation
- 28:10Cooperative Policy
- 29:42HighLevel Strategies
related talks
45mAGT and Machine Learning
31mDecision theoretic foundations for human-AI collaboration
42mRethinking AI Alignment and Evaluation under Heterogeneous Preferences
19mTraining Agentic Reasoners
38mRetrospective on multi-agent learning as a path to intelligence
45mPanel Discussion
31mFrom Single AI to Multi AI Systems
44mAI Ecosystems: Structure, Strategy, and Regulation