Publication
Une nouvelle approche au General Game Playing dirigée par les contraintes
Summary
This thesis develops a new constraint-based approach to General Game Playing (GGP), with the goal of designing a program able to play a wide range of strategic games from their formal rules alone.
The core idea is to translate games described in Game Description Language (GDL) and its extension GDL-II into stochastic constraint satisfaction problems (SCSPs). This provides a dense and declarative representation of games, supports the modelling of strategies, and makes it possible to exploit general-purpose constraint reasoning.
The thesis then introduces a family of algorithms including MAC-UCB, MAC-UCB-II, and later symmetry-aware variants, combining constraint propagation with Monte Carlo sampling in order to play both complete-information and incomplete-information games.
It also presents WoodStock, the general game player implementing this approach, and shows experimentally that constraint-based techniques can compete with and surpass state-of-the-art General Game Playing methods.
Distinctions and impact
This thesis received the AFIA PhD Thesis Prize in Artificial Intelligence (ex-aequo, 2017) .
It also underpins the development of WoodStock, which won the International General Game Playing Competition (IGGPC) 2016 .
The work was also highlighted in a CNRS INS2I news article: « Une intelligence artificielle dirigée par les contraintes championne de General Game » .
Full reference
Piette, É. (2016). Une nouvelle approche au General Game Playing dirigée par les contraintes. PhD thesis, Université d'Artois, CRIL — Centre de Recherche en Informatique de Lens, soutenue le 9 décembre 2016.
BibTeX
@phdthesis{piette2016ggp,
author = {Piette, {\'E}ric},
title = {Une nouvelle approche au General Game Playing dirig{\'e}e par les contraintes},
school = {Universit{\'e} d'Artois},
year = {2016},
month = dec,
note = {CRIL -- Centre de Recherche en Informatique de Lens, soutenue le 9 d{\'e}cembre 2016}
}Context
This thesis is the central document bringing together the different strands of the early part of my research: General Game Playing, stochastic constraint programming, imperfect information, Monte Carlo methods, and symmetry detection.
It also provides the scientific foundation for the WoodStock player and for a broader research direction centred on general, strategic, and human-relevant AI systems.