A comparison of feature functions for Tetris strategies

Finding optimal strategies for the game of Tetris is an interesting NP-complete problem that has attracted several AI researchers. Their approaches display subtle variations in the implementation details, with unclear relationships between these details and Tetris performance. This, combined with the absence of confidence intervals in most published results, makes the evaluation and comparison of Tetris strategies and optimization methods very difficult.To look furhter into this unclear relationship, we would re-create every environnment described in several publications. An evolutionary algorithm was executed within each environment to create an AI and their performance compared against each other. The scores differed \textit{substantially}. This suggests that some aspects of the Tetris environment greatly affects the potentially obtainable performance of an AI. We come to the unfortunate conclusion that nearly no results of existing publications can be used to compare optimization methods against each other in terms of suitability for Tetris due to this reason