Playing (Almost-)Optimally in Concurrent B\"uchi and co-B\"uchi Games

TL;DR

This paper investigates the existence and complexity of optimal strategies in concurrent stochastic games with B"uchi and co-B"uchi objectives, revealing conditions for positionality and the necessity of memory.

Contribution

It provides characterizations of local interactions that guarantee positional (almost-)optimal strategies in concurrent B"uchi and co-B"uchi games.

Findings

Existence of optimal strategies varies with objectives and game forms.

Characterizations of local interactions ensure positionality of strategies.

Infinite memory may be required for some objectives.

Abstract

We study two-player concurrent stochastic games on finite graphs, with B\"uchi and co-B\"uchi objectives. The goal of the first player is to maximize the probability of satisfying the given objective. Following Martin's determinacy theorem for Blackwell games, we know that such games have a value. Natural questions are then: does there exist an optimal strategy, that is, a strategy achieving the value of the game? what is the memory required for playing (almost-)optimally? The situation is rather simple to describe for turn-based games, where positional pure strategies suffice to play optimally in games with parity objectives. Concurrency makes the situation intricate and heterogeneous. For most {\omega}-regular objectives, there do indeed not exist optimal strategies in general. For some objectives (that we will mention), infinite memory might also be required for playing optimally or almost-optimally. We also provide characterizations of local interactions of the players to ensure positionality of (almost-)optimal strategies for B\"uchi and co-B\"uchi objectives. This characterization relies on properties of game forms underpinning the formalism for defining local interactions of the two players. These well-behaved game forms are like elementary bricks which, when they behave well in isolation, can be assembled in graph games and ensure the good property for the whole game.