New summary measures and datasets for the multi-project scheduling problem

In recent years, more researchers have devoted their attention to the resource-constrained multi-project scheduling problem, resulting in a growing body of knowledge on solution procedures. A key factor in the comparison of these procedures is the availability of benchmark datasets that cover a large part of the feature space. Otherwise, one risks that the conclusions from experiments on these sets do not hold when they are repeated on a different set. In this paper we propose new multi-project datasets that contain instances with a wide variety of characteristics. We first develop several new summary measures that describe three types of portfolio characteristics, two of the three types are not present in any of the existing datasets. Second, an algorithm is developed that can generate instances with the desired parameter values in a controlled manner. With this procedure, we create three datasets that each focus on one of the characteristics and a fourth dataset that contains all combinations. The computational results show (a) that these sets cover a significantly larger part of the feature space than existing benchmark libraries and (b) that they are more challenging for advanced algorithms.