Approximating the Termination Value of One-Counter MDPs and Stochastic Games

TL;DR

This paper proves that the quantitative termination values of one-counter MDPs and stochastic games are computable, enabling approximation within any desired precision and the computation of epsilon-optimal strategies.

Contribution

It establishes the computability of approximation problems for termination values in OC-MDPs and OC-SSGs, a previously open question, and introduces martingale-based techniques for this purpose.

Findings

All quantitative approximation problems are computable.

Epsilon-optimal strategies can be computed for both players.

A novel martingale approach is used to derive bounds.

Abstract

One-counter MDPs (OC-MDPs) and one-counter simple stochastic games (OC-SSGs) are 1-player, and 2-player turn-based zero-sum, stochastic games played on the transition graph of classic one-counter automata (equivalently, pushdown automata with a 1-letter stack alphabet). A key objective for the analysis and verification of these games is the termination objective, where the players aim to maximize (minimize, respectively) the probability of hitting counter value 0, starting at a given control state and given counter value. Recently, we studied qualitative decision problems ("is the optimal termination value = 1?") for OC-MDPs (and OC-SSGs) and showed them to be decidable in P-time (in NP and coNP, respectively). However, quantitative decision and approximation problems ("is the optimal termination value ? p", or "approximate the termination value within epsilon") are far more challenging. This is so in part because optimal strategies may not exist, and because even when they do exist they can have a highly non-trivial structure. It thus remained open even whether any of these quantitative termination problems are computable. In this paper we show that all quantitative approximation problems for the termination value for OC-MDPs and OC-SSGs are computable. Specifically, given a OC-SSG, and given epsilon > 0, we can compute a value v that approximates the value of the OC-SSG termination game within additive error epsilon, and furthermore we can compute epsilon-optimal strategies for both players in the game. A key ingredient in our proofs is a subtle martingale, derived from solving certain LPs that we can associate with a maximizing OC-MDP. An application of Azuma's inequality on these martingales yields a computable bound for the "wealth" at which a "rich person's strategy" becomes epsilon-optimal for OC-MDPs.