Characterizing Locality in Decoder-Based EAs for the Multidimensional Knapsack Problem