Automated puzzle difficulty estimation

Abstract

We introduce a method for automatically rating the difficulty of puzzle game levels. Our method takes multiple aspects of the levels of these games, such as level size, and combines these into a difficulty function. It can simply be adapted to most puzzle games, and we test it on three different ones: Flow, Lazors and Move. We conducted a user study to discover how difficult players find the levels of a set and use this data to train the difficulty function to match the user-provided ratings. Our experiments show that the difficulty function is capable of rating levels with an average error of approximately one point in Lazors and Move, and less than half a point in Flow, on a difficulty scale of 1-10.