The Complexity of Infinite-Horizon General-Sum Stochastic Games: Turn-Based and Simultaneous Play
summary
Shows non‑stationary Nash equilibria computable poly(S,A) and stationary Nash equilibria computable poly(A), but proves computing an approximate stationary NE in turn‑based infinite‑horizon general‑sum stochastic games is PPAD‑complete in S for sufficiently large discount (gamma > 1/2). Key tools: a pseudo‑linear single‑state deviation property to get PPAD membership, and a reduction from generalized arithmetic circuits using gadgets with long output paths to force locality; hard instances use N players with rewards at ≤4 states or a 6‑player encoding.
AI-written summary · about
chapters (21)
- 0:00Intro
- 1:20General-sum stochastic (Markov) games
- 3:36Turn-based stochastic games (TBSG's)
- 6:29Possible evidence for polynomial-time computability
- 8:23Possible evidence for intractability
- 9:27Key questions: The complexity of stationary equilibria in infinite-horizon general-sum stochastic games
- 10:24Outline of talk
- 11:34Infinite-horizon simultaneous stochastic games (SimSG)
- 14:35Infinite-horizon turn-based stochastic games (TBSG)
- 15:39The PPAD complexity class
- 16:43Challenges in studying SimSG and TBSG
- 21:04Structural properties of SimSG and TBSG: Pseudolinearity
- 25:10Building blocks for pseudo-linearity: "Two-policy" MDP "Two-policy" MDP: one state has two actions, every other state has only one action
- 27:19Progress so far: The complexity of stationary equilibria in infinite-horizon general-sum stochastic games
- 34:09A class of problems known to be PPAD-hard The generalized arithmetic circuit problem (G Circuit Input: A circuit that may contains cycles) of real-valued arithmetic operations, "fan-out"1
- 36:17Review: Graphical "game gadget" (for equality gate)
- 39:16TBSG game gadget for equality gate?
- 41:28TBSG game gadgets with induced locality and linearity
- 42:21Omitted details of reduction
- 42:49Summary of our complexity characterization
- 44:27Open problems Other instances of general-sum structure for polynomial-time stationary NE computation
