In the context of real-valued evolutionary optimization in high dimensional domains, understanding and exploiting the problem structure can lead to significant improvements in final result quality while also lowering the computational burdens by cutting down evaluation time. This dissertation presents novel approaches for linkage learning and gene sensitivity detection through machine learning methods in the real-valued domains and a proposed idea to jointly represent these measures. A surrogate-assisted perturbation-check for non-linearity that does not stress the true fitness function is introduced and various machine-learning methods are employed and compared in terms of their ability to rank gene importance. Furthermore, novel surrogate assisted crossover operators that incorporate linkage knowledge through crossover masks are defined and evaluated on synthetic fitness functions to empirically validate their utility. Finally, a new benchmark with overlapping linkage groups of increasing size is presented, which provides a platform for comparison of real-valued global optimization algorithms.