| Abstract: |
In the global economy, billions of dollars of merchandise are routed using software that, at its core, uses optimization technology. Over many decades, researchers have devised different approaches to make algorithms faster, and this is true for Benders decomposition as well. Benders speeds up finding an optimal solution to a problem with millions of variables and constraints by iteratively learning which constraints are important and considering only these constraints. Our idea is that selectively choosing the constraints that eliminate the largest number of irrelevant solutions at each step would lead to finding the optimal solution in the fewest number of Benders steps. Geometrically, this amounts to choosing so-called deep cuts. Of course, in attempting to minimize the number of steps, we do need to spend more time taking each individual step, but our experimental results on several types of problems arising in supply chain analytics show that this approach makes sense and significantly reduces the solution time. Since its inception, Benders decomposition (BD) has been successfully applied to a wide range of large-scale mixed-integer (linear) problems. The key element of BD is the derivation of Benders cuts, which are often not unique. In this paper, we introduce a novel unifying Benders cut selection technique based on a geometric interpretation of cut depth, produce deepest Benders cuts based on ℓp -norms, and study their properties. Specifically, we show that deepest cuts resolve infeasibility through minimal deviation (in a distance sense) from the incumbent point, are relatively sparse, and may produce optimality cuts even when classic Benders would require a feasibility cut. Leveraging the duality between separation and projection, we develop a guided projections algorithm for producing deepest cuts, exploiting the combinatorial structure and decomposability of problem instances. We then propose a generalization of our Benders separation problem, which not only brings several well-known cut selection strategies under one umbrella, but also, when endowed with a homogeneous function, enjoys several properties of geometric separation problems. We show that, when the homogeneous function is linear, the separation problem takes the form of the minimal infeasible subsystems (MIS) problem. As such, we provide systematic ways of selecting the normalization coefficients of the MIS method and introduce a directed depth-maximizing algorithm for deriving these cuts. Inspired by the geometric interpretation of distance-based cuts and the repetitive nature of two-stage stochastic programs, we introduce a tailored algorithm to further facilitate deriving these cuts. Our computational experiments on various benchmark problems illustrate effectiveness of deepest cuts in reducing both computation time and number of Benders iterations and producing high-quality bounds at early iterations. [ABSTRACT FROM AUTHOR] |
