Multilevel Reactive Tabu Search for Graph Partitioning

In this paper, new multilevel heuristics are proposed for nding a balanced bipartition of a graph. Multilevel schemes have been recently introduced for quickly solving the problem on graphs with thousands of nodes, graphs often present in domain-decomposition techniques for parallel computing. These heuristics rstly reduce the original graph by collapsing its nodes and arcs, thus generating a hierarchy of smaller and smaller graphs, then nd a good partition of the smallest graph, and nally project and reene the so found partition backwards on the graph hierarchy. The multilevel algorithms proposed in this paper employ a novel greedy method and tabu search techniques during the partitioning and uncoarsening phases. Experimental results on a wide collection of benchmark graphs show that partitions of high quality are found in small computing times. Moreover, the eeect of various parameter choices during the partitioning phase is experimentally studied.