The Detrimentality of Crossover

A survey of techniques for characterising fitness landscapes and some possible ways forward

“…In complex problems, the interactions between variables can take many different forms. Studies have shown that linearly separable functions are easier for genetic algorithms to solve than non-linearly separable functions [63,13]. Naudts and Naudts [49] argue that it is the type of interaction (functions with first and second order dependencies) rather than the amount of higher order interaction that influences the difficulty of the problem for search algorithms.…”

“…Each factor is clearly important to some degree and it opens the question: Are there characteristics which are also important, but are missing from the list of available techniques? For example, symmetry is known to influence problem difficulty [63,85,49,13], but to the authors' knowledge there are no known techniques for measuring symmetry in fitness landscapes. Another example is the degree of variable interdependency.…”

“…This could be due to the phenomenon where two dissimilar good (symmetrical) solutions, crossed over, result in inferior children. Other studies have shown that genetic algorithms show improved performance on landscapes with axial biases [13] and that a rotation of the coordinate system for such problems (resulting in the loss of symmetry) causes severe algorithmic performance loss [63].…”