Using Genetic Algorithms for Investigating Specific Regions of the Solution Space
Irfan Younas, Rassul Ayani, Johan Schubert · 2011
Abstract—In many practical cases decision makers are interested to understand the whole solution space, including possible outliers. By outlier we mean there is a solution that is theoretically possible, though with very low probability that it occurs. In many combinatorial problems, this is a very challenging task. During the past decade, the Data Farming community has done substantial work on developing methods and techniques for better understanding of the solution space. The data farming community has also looked at the design of experiments and used Latin hypercube (LH) techniques for this purpose. The LH is proven to be one of the important sampling methods for selecting a representative subset of the input space. In this paper, we consider a company X that wants to outsource m subprojects of a given project P. We assume that there are n potential subcontractors for each subproject. Thus, there will be n m ways to assign the subprojects to the potential subcontractors. The project manager is interested to find those assignments that complete the project within a given time and a given cost frame. An exhaustive examination of all assignments is not feasible, if m and n are big numbers. We propose an objective-based genetic algorithm (GA) for finding the set of assignments that are mapped onto a given subset of the solution space. It means, as opposed to the design of experiment techniques, we start from the solution space and try to find the combinations of the input parameter values that can lead to a specific region of the solution space. By some numerical examples, we show how our GA identifies the set of such feasible assignments.