Learning Decision Trees for the Analysis of Optimization Heuristics

Decision trees are widely used to represent information extracted from data sets. In studies on heuristics for optimization, there are two types of information in which we may be interested: how the parameters of the algorithm affect its performance and which characteristics of the instances determine a difference in the performance of the algorithms. Tree-based learning algorithms, as they exist in several software packages, do not allow to model thoroughly experimental designs for answering these types of questions. We try to overcome this issue and devise a new learning algorithm for the specific settings of analysis of optimization heuristics.