An analytical comparison of optimization problem generation methodologies

Heuristics are an increasingly popular solution method for combinatorial optimization problems. Heuristic use often frees the modeler from some of the restrictions placed on classical optimization methods required to constrain problem complexity. As a result, modelers are using heuristics to tackle problems previously considered unsolvable, improve performance over classical optimization methods, and open new avenues of empirical study. Researchers should fully understand key test problem attributes and sources of variation to produce efficient and effective optimization studies. These problem attributes and sources of variation are reviewed. Problem correlation structure significantly effects algorithm performance but is often overlooked or ignored in empirical studies. The paper analyzes the correlation structure among a set of standard multidimensional knapsack problems and recommends an improved approach to synthetic, or randomly generated optimization problems for the empirical study of solution algorithms for combinatorial optimization problems.