Skip to Main content Skip to Navigation
Conference papers

Apprendre à jouer aux jeux à deux joueurs à information parfaite sans connaissance

Quentin Cohen-Solal 1
1 Equipe MAD - Laboratoire GREYC - UMR6072
GREYC - Groupe de Recherche en Informatique, Image, Automatique et Instrumentation de Caen
Abstract : In this paper, several techniques for learning game states evaluation functions by reinforcement are proposed. The first is to learn the values of the game tree instead of restricting oneself to the value of the root. The second is to replace the classic gain of a game (+1 / −1) with a heuris-tic favoring quick wins and slow defeats. The third corrects some evaluation functions taking into account the resolution of states. The fourth is a new action selection distribution. Finally, the fifth is a modification of the minimax with unbounded depth extending the best sequences of actions to the terminal states. In addition, we propose another variant of the unbounded minimax, which plays the safest action instead of playing the best action. The experiments conducted suggest that this improves the level of play during confrontations. Finally, we apply these different techniques to design a program-player to the Hex game (size 11) reaching the level of Mohex 2.0 with reinforcement learning from self-play without knowledge.
Document type :
Conference papers
Complete list of metadatas

Cited literature [32 references]  Display  Hide  Download
Contributor : Jérôme Lang <>
Submitted on : Wednesday, October 23, 2019 - 12:18:47 PM
Last modification on : Saturday, October 26, 2019 - 2:06:54 AM
Long-term archiving on: : Friday, January 24, 2020 - 5:37:01 PM


paper 6.pdf
Files produced by the author(s)


  • HAL Id : hal-02328750, version 1


Quentin Cohen-Solal. Apprendre à jouer aux jeux à deux joueurs à information parfaite sans connaissance. Conférence Nationale en Intelligence Artificielle, Jul 2019, Toulouse, France. ⟨hal-02328750⟩



Record views


Files downloads