Evolutionary k-way Node Separators

Small node separators for large graphs are used in a variety of ways, from divide-andconquer algorithms to efficient route planning algorithms. We present a new algorithm for finding small k-way node separators on connected undirected graphs. We use an evolutionary algorithm to combine different separators in order to generate improved offsprings with locally good parts of both parent separators. Additionally, we also propose a new k-way local search to further improve already existing separators. Our experimental evaluation of our algorithm shows that we have an average improvement of 18% and 23% respectively for our two different configurations and a maximum improvement of 64% compared to the already existing algorithm in KaHIP for finding k-way node separators.