Stochastic Games with Limited Public Memory

TL;DR

This paper proves that near-optimal strategies in two-player zero-sum stochastic games can be implemented with logarithmic memory, significantly improving previous bounds, but such strategies cannot be achieved with bounded public memory.

Contribution

It establishes that uniform ε-optimal strategies require only O(log n) memory states, and demonstrates the limitations of bounded public memory strategies in such games.

Findings

Near-optimal strategies use at most O(log n) memory states.

Time-independent strategies can achieve near-optimality with high probability.

Bounded public memory strategies cannot guarantee near-optimal payoffs in the Big Match.

Abstract

We study the memory resources required for near-optimal play in two-player zero-sum stochastic games with the long-run average payoff. Although optimal strategies may not exist in such games, near-optimal strategies always do. Mertens and Neyman (1981) proved that in any stochastic game, for any $\varepsilon>0$, there exist uniform $\varepsilon$-optimal memory-based strategies -- i.e., strategies that are $\varepsilon$-optimal in all sufficiently long $n$-stage games -- that use at most $O(n)$ memory states within the first $n$ stages. We improve this bound on the number of memory states by proving that in any stochastic game, for any $\varepsilon>0$, there exist uniform $\varepsilon$-optimal memory-based strategies that use at most $O(\log n)$ memory states in the first $n$ stages. Moreover, we establish the existence of uniform $\varepsilon$-optimal memory-based strategies whose memory updating and action selection are time-independent and such that, with probability close to 1, for all $n$, the number of memory states used up to stage $n$ is at most $O(\log n)$. This result cannot be extended to strategies with bounded public memory -- even if time-dependent memory updating and action selection are allowed. This impossibility is illustrated in the Big Match -- a well-known stochastic game where the stage payoffs to Player 1 are 0 or 1. Although for any $\varepsilon > 0$, there exist strategies of Player 1 that guarantee a payoff {exceeding} $1/2 - \varepsilon$ in all sufficiently long $n$-stage games, we show that any strategy of Player 1 that uses a finite public memory fails to guarantee a payoff greater than $\varepsilon$ in any sufficiently long $n$-stage game.