Parallel MO-PBIL: Computing pareto optimal frontiers efficiently with applications in reinsurance analytics

In this paper we propose MO-PBIL, a parallel multidimensional variant of the Population Based Incremental Learning (PBIL) technique that executes efficiently on both multi-core and many-core architectures. We show how MO-PBIL can be used to address an important problem in Reinsurance Risk Analytics namely the Reinsurance Contract Optimization problem. A mix of vectorization and multithreaded parallelism is used to accelerate the three main computational steps: objective function evaluation, multidimensional dominance calculations, and multidimensional clustering. Using MO-PBIL, reinsurance contract optimization problems with a 5% discretization and 7 or less contractual layers (subcontracts) can be solved in under a 1 minute on a single workstation or server. Problems with up to 15 layers, which previously took a month or more of computation to solve, can now be solved in less than 10 minutes.