Résumé : Most research about the game of Minesweeper has focussed on inferring which cells may or may not contain a mine, discussing the complexity of this problem (and its variants) and proposing efficient resolution techniques. While this inference task is indeed crucial to playing Minesweeper, this paper comes back to the original game as a whole, modelling it as a sequential decision-making problem ---more precisely as a Partially Observable Markov Decision Process (POMDP)--- before proposing and studying various heuristics to solve the key problem of selecting the next cell to play.