Parallelization strategies for rapid and robust evolutionary multiobjective optimization in water resources applications

This study uses a formal metrics-based framework to demonstrate the Master–Slave (MS) and the Multiple-Population (MP) parallelization schemes for the Epsilon-Nondominated Sorted Genetic Algorithm-II (ε-NSGAII). The MS and MP versions of the ε-NSGAII generalize the algorithm’s auto-adaptive population sizing, ε-dominance archiving, and time continuation to a distributed processor environment using the Message Passing Interface. This study uses three test cases to compare the MS and MP versions of the ε-NSGAII: (1) an extremely difficult benchmark test function termed DTLZ6 drawn from the computer science literature, (2) an unconstrained, continuous hydrologic model calibration test case for the Leaf River near Collins, Mississippi, and (3) a discrete, constrained four-objective long-term groundwater monitoring (LTM) application. The MP version of the ε-NSGAII is more effective than the MS scheme when solving DTLZ6. Both the Leaf River and the LTM test cases proved to be more appropriately suited to the MS version of the ε-NSGAII. Overall, the MS version of the ε-NSGAII exhibits superior performance on both of the water resources applications, especially when considering its simplicity and ease-of-implementation relative to the MP scheme. A key conclusion of this study is that a simple MS parallelization strategy can exploit time-continuation and parallel speedups to dramatically improve the efficiency and reliability of evolutionary multiobjective algorithms in water resources applications.