Subset Feature Selection problems can have several attributes which may make Messy Genetic Algorithms an appropriate optimization method. First, competitive solutions may often use only a small percentage of the total available features; this can not only o er an advantage to Messy Genetic Algorithms, it may also cause problems for other types of evolutionary algorithms. Second, the evaluation of small blocks of features is naturally decomposable. Thus, there is no di culty evaluating underspeci ed strings. We apply variants of the Messy Genetic Algorithm to a application in computer vision with very good results. We also apply variants of the Fast Messy Genetic Algorithm to synthethic test problems.