Sébastien Tabary’s research while affiliated with University of Lille Nord de France and other places

Symmetry detection is a promising approach for reducing the search tree of games. In General Game Playing (GGP), where any game is compactly represented by a set of rules in the Game Description Language (GDL), the state-of-the-art methods for symmetry detection rely on a rule graph associated with the GDL description of the game. Though such rule-based symmetry detection methods can be applied to various tree search algorithms, they cover only a limited number of symmetries which are apparent in the GDL description. In this paper, we develop an alternative approach to symmetry detection in stochastic games that exploits constraint programming techniques. The minimax optimization problem in a GDL game is cast as a stochastic constraint satisfaction problem (SCSP), which can be viewed as a sequence of one-stage SCSPs. Minimax symmetries are inferred according to the microstructure complement of these one-stage constraint networks. Based on a theoretical analysis of this approach, we experimentally show on various games that the recent stochastic constraint solver MACUCB, coupled with constraint-based symmetry detection, significantly outperforms the standard Monte Carlo Tree Search algorithms, coupled with rule-based symmetry detection. This constraint-driven approach is also validated by the excellent results obtained by our player during the last GGP competition.

A portfolio SAT solver has to share clauses in order to be efficient. In a distributed environment, such sharing implies additional problems: more information has to be exchanged and communications among solvers can be time consuming. In this paper, we propose a new version of the state-of-the-art SAT solver Syrup that is now able to run on distributed architectures. We analyze and compare different programming models of communication. We show that, using a dedicated approach, it is possible to share many clauses without penalizing the solvers. Experiments conducted on SAT 2016 benchmarks with up to 256 cores show that our solver is very effective and outperforms other approaches. This opens a broad range of possibilities to boost parallel solvers needing to share many data.

This article describes WoodStock, the first general game player modeling each game from the General Game Playing (GGP) by a stochastic constraint network (SCSP). Each action played is decided by the resolution of this last one by the algorithm MAC-UCB. After the translation of an instance described in Game Description Language (GDL) in a network representative of the state of the game at any time, WoodStock solves each state by the maintening arc-consistency algorithm (MAC) iteratively guided by the bandit-based stochastic sampling (UCB) of the next states. Thanks to this algorithm, WoodStock is since march 2016, the leader of the GGP Tiltyard continuous tournament. Moreover, in its last version exploiting the game symmetries finding by the constraint symmetry detection, the search space associated with a game is significatively reduced. With that, WoodStock is now the GGP champion after its victory at the International General Game Playing Competition 2016 (IGGPC 2016) organized by the Stanford University.

We present and evaluate AmPharoS, a new parallel SAT solver based on the divide and conquer paradigm. This solver, designed to work on a great number of cores, runs workers on sub-formulas restricted to cubes. In addition to classical clause sharing, it also exchange extra information associated to the cubes. Furthermore, we propose a new criterion to dynamically adapt both the amount of shared clauses and the number of cubes. Experiments show that, in general, AmPharoS correctly adjusts its strategy to the nature of the problem. Thus, we show that our new parallel approach works well and opens a broad range of possibilities to boost parallel SAT solver performances.

The challenge of General Game Playing (GGP) is to devise game playing programs that take as input the rules of any strategic game, described in the Game Description Language (GDL), and that effectively play without human intervention. The aim of this paper is to address the GGP challenge by casting GDL games (potentially with chance events) into the Stochastic Constraint Satisfaction Problem (SCSP). The stochastic constraint network of a game is decomposed into a sequence of µ SCSPs (also know as one-stage SCSP), each associated with a game round. Winning strategies are searched by coupling the MAC (Maintaining Arc Consistency) algorithm, used to solve each µ SCSP in turn, together with the UCB (Upper Confidence Bound) policy for approximating the values of those strategies obtained by the last µ SCSP in the sequence. Extensive experiments conducted on various GDL games with different deliberation times per round, demonstrate that the MAC-UCB algorithm significantly outperforms the state-of-the-art UCT (Upper Confidence bounds for Trees) algorithm.

The game description language with incomplete information (GDL-II) is expressive enough to capture partially observable stochastic multi-agent games. Unfortunately, such expressiveness does not come without a price: the problem of finding a winning strategy is NE XP NP -hard, a complexity class which is far beyond the reach of modern constraint solvers. In this paper, we identify a P SPACE -complete fragment of GDL-II, where agents share the same (partial) observations. We show that this fragment can be cast as a decomposable stochastic constraint satisfaction problem (SCSP) which, in turn, can be solved using general-purpose constraint programming techniques. Namely, we develop a constraint-based sequential decision algorithm for GDL-II games which exploits constraint propagation and Monte Carlo sampling based. Our algorithm, validated on a wide variety of games, significantly outperforms the state-of-the-art general game playing algorithms.

1 The aim of General Game Playing (GGP) is to devise game playing algorithms which are not dedicated to a particular strategic game, but are general enough to effectively play a wide variety of games. A tournament is held every year by AAAI, in which artificial game players are supplied the rules of arbitrary new games and, without human intervention, have to play these game optimally. Games rules are described in a declarative representation language, called GDL for Game Description Language. The lastest version of this language is expressive enough to describe finite multi-player games with uncertain and incomplete information. GGP algorithms include, among others, answer set programming methods, automated construction of evaluation functions, and Monte-Carlo methods such as Upper Confidence bounds for Trees (UCT). Beyond its play value, GGP offers a rigorous setting for modeling and analyzing sequential decision-making algorithms in multi-agent environments. By providing a declarative approach for representing and solving combinatorial problems, Constraint Programming appears as a promising technology to address the GGP challenge. Currently, several constraint-based formalisms have already been proposed to model and solve games; they include notably Quantified CSP, Strategic CSP and Constraint Games. Most of these formalisms are, however, restricted to deterministic, perfect information games: during each round of the game, players have full access to the current state and their actions have determin-istic effects. This paper focuses on stochastic games, with chance events, using the framework of stochastic constraint networks. More precisely, we study a fragment of the Stochastic Constraint Satisfaction Problem (SCSP), that captures GDL games with uncertain (but complete) information. Interestingly, the SCSP for this class of games can be decomposed into a sequence of µSCSPs (a.k.a one-stage stochastic constraint satisfaction problems). Based on this decomposition, we propose a sequential decision-making algorithm, MAC-UCB, that combines the MAC algorithm (Maintaining Arc Consistency) for solving each µSCSP, and the multi-armed bandits Upper Confidence Bound (UCB) method for approximating the expected utility of strategies. We show that in practice MAC-UCB significantly outperforms (the multi-player version of) UCT, which is the state-of-the-art GGP algorithm for stochastic games. MAC-UCB also dominates FC-UCB, a variant where MAC is replaced with the Forward Checking (FC) algorithm. Such conclusions are drawn from comparing the performance of these algorithms, using extensive experiments (about 1, 350, 000 face-offs) over a wide range of GDL games. 1 This paper has been published in Constraints 21(1), the Journal Fast Track issue of CP'15.