Parallel Evolutionary Multi-Objective Optimization on Large, Heterogeneous Clusters: An Applications Perspective

Abstract

Real-world operational use of parallel multi-objective evolutionary algorithms requires successful searches in constrained wall-clock periods, limited trial-and-error algorithmic analysis, and scalable use of heterogeneous computing hardware. This study provides a cross-disciplinary collaborative effort to assess and adapt parallel multi-objective evolutionary algorithms for operational use in satellite constellation design using large dedicated clusters with heterogeneous processor speeds/architectures. A statistical, metric-based evaluation framework is used to demonstrate how time-continuation, asynchronous evolution, dynamic population sizing, and epsilon dominance archiving can be used to enhance both simple master-slave parallelization strategies and more complex multiple-population schemes. Results for a benchmark constellation design coverage problem show that simple masterslave schemes that exploit time-continuation are often sufficient and potentially superior to complex multiple-population schemes.

Many-objective reconfiguration of operational satellite constellations with the Large-Cluster Epsilon Non-dominated Sorting Genetic Algorithm-II

“…Table 1 provides an overview of the objective vector, x f , for reference. [20] is used with the original Epsilon Nondominated Sorting Genetic Algorithm-II ( -NSGA-2) [21] and adapted for large heterogeneous clusters (LC--NSGA-2) [9]. The -NSGA-2 provides three primary innovations (epsilon-dominance archiving [22], auto-adaptive population sizing [23], use of time continuation [24]) that address some of the difficulties previously experienced using its parent algorithm, the NSGA-2, for solving problems in the constellation design domain [11].…”

“…The second, and key enabler towards eliminating the scalability, is the development of asynchronous evolution [9] wherein the master core does not wait for the final objective vector evaluations of a given population before proceeding to the next generation.…”

“…As soon as the 'To-Evaluate' list has sent its last chromosome to a slave core, those individuals residing in the 'Finished' list form a variable length child population, N v , that is immediately merged with the parent population N. Even though the child population size varies, the total population size reaches a steady-state where the number of individuals left behind approximately equals the number of individuals from earlier generations where the slave cores have finished processing. The net result of asynchronous evolution is near 100% core utilization for clusters containing on the order of hundreds of nodes [9]. While the scalability issue is resolved with LC--NSGA-II, another modification was required to cope with potential extendedduration processing time and system failures.…”

“…Recent work [9]- [11] has demonstrated that parallel Multi-Objective Evolutionary Algorithms (pMOEAs) provide an efficient and effective means for the design of satellite constellations optimized for two-dimensional performance tradeoffs on geometric coverage. The reconfiguration of satellite constellations however, requires that a broader range of cost and risk objectives be considered along with performance.…”

“…In this work, the Large-Cluster Epsilon Non-dominated Sorting Genetic Algorithm-II (LC--NSGA-2) [9] is used to approximate Pareto-hypervolumes for the reconfiguration problem. The framework is demonstrated on loss scenarios for the Global Positioning System (GPS) constellation and a a posteriori decision support methodology is presented for sorting through the potentially thousands of non-dominated designs for down-selection to final decisions.…”