David Hadka

Borg: An auto-adaptive many-objective evolutionary computing framework

This study introduces the Borg multi-objective evolutionary algorithm (MOEA) for many-objective, multimodal optimization. The Borg MOEA combines -dominance, a measure of convergence speed named -progress, randomized restarts, and auto-adaptive multioperator recombination into a unified optimization framework. A comparative study on 33 instances of 18 test problems from the DTLZ, WFG, and CEC 2009 test suites demonstrates Borg meets or exceeds six state of the art MOEAs on the majority of the tested problems. The performance for each test problem is evaluated using a 1,000 point Latin hypercube sampling of each algorithms feasible parameteri- zation space. The statistical performance of every sampled MOEA parameterization is evaluated using 50 replicate random seed trials. The Borg MOEA is not a single algorithm; instead it represents a class of algorithms whose operators are adaptively selected based on the problem. The adaptive discovery of key operators is of particular importance for benchmarking how variation operators enhance search for complex many-objective problems.

Diagnostic assessment of search controls and failure modes in many-objective evolutionary optimization

The growing popularity of multiobjective evolutionary algorithms (MOEAs) for solving many-objective problems warrants the careful investigation of their search controls and failure modes. This study contributes a new diagnostic assessment framework for rigorously evaluating the effectiveness, reliability, efficiency, and controllability of MOEAs as well as identifying their search controls and failure modes. The framework is demonstrated using the recently introduced Borg MOEA, -NSGA-II, -MOEA, IBEA, OMOPSO, GDE3, MOEA/D, SPEA2, and NSGA-II on 33 instances of 18 test problems from the DTLZ, WFG, and CEC 2009 test suites. The diagnostic framework exploits Sobols variance decomposition to provide guidance on the algorithms’ non-separable, multi-parameter controls when performing a many-objective search. This study represents one of the most comprehensive empirical assessments of MOEAs ever completed.

Large-scale parallelization of the Borg multiobjective evolutionary algorithm to enhance the management of complex environmental systems

The Borg MOEA is a self-adaptive multiobjective evolutionary algorithm capable of solving complex, many-objective environmental systems problems efficiently and reliably. Water and environmental resources problems pose significant computational challenges due to their potential for large Pareto optimal sets, the presence of disjoint Pareto-optimal regions that arise from discrete choices, multi-modal suboptimal regions, and expensive objective function calculations. This work develops two large-scale parallel implementations of the Borg MOEA, the master-slave and multi-master Borg MOEA, and applies them to a highly challenging risk-based water supply portfolio planning problem. The performance and scalability of both implementations are compared on up to 16384 processors. The multi-master Borg MOEA is shown to scale efficiently on tens of thousands of cores while dramatically improving the reliability of attaining high-quality solutions. Our results dramatically expand the scale and scope of complex environmental systems that can be addressed using many-objective evolutionary optimization. Massively parallel extensions of the Borg multiobjective evolutionary algorithm.Cooperating instances of master slave parallelizations dramatically enhance search.Parallelizing the Borg MOEA improves efficiency, search quality, and reliability.Discrete event simulation shows theoretical scalability for 100,000 compute cores.250+ years of search possible in 24?h.

Evolving many‐objective water management to exploit exascale computing

This study contributes one of the largest parallel scalability experiments ever attempted within the water resources literature to date, encompassing 2000 years of computational time. A severely challenging multiobjective benchmark problem focusing on urban water portfolio planning under uncertainty in the Lower Rio Grande Valley (LRGV) is used to demonstrate that a multimaster variant of the Borg multiobjective evolutionary algorithm (MOEA) can be used efficiently on more than 524,288 compute cores. The scalability of the multimaster Borg MOEA enables users to compress up to 20 years of computational work into 20 min of actual wall-clock time. Beyond these temporal efficiency gains, metric-based statistical assessments of solution quality show that the multimaster Borg MOEA dramatically enhances the effectiveness and reliability of the algorithms autoadaptive search features. Theoretical algorithmic analysis shows that the multimaster Borg MOEA could maintain high levels of parallel scalability on future exascale computing platforms (i.e., millions of compute cores). These results mark a fundamental expansion of the scope, computational demands, and difficulties that can be addressed in multiobjective water resources applications.

Scalability Analysis of the Asynchronous, Master-Slave Borg Multiobjective Evolutionary Algorithm

The Borg Multiobjective Evolutionary Algorithm (MOEA) is a new, efficient, and robust optimizer that outperforms competing optimization methods on numerous complex engineering problems. To date, the Borg MOEA has been successfully applied to problems ranging from aerospace applications to water resources engineering. Problems from these domains often involve expensive design evaluations that require large-scale parallel algorithms to produce results in a reasonable amount of time. This study presents the first theoretical and experimental look at parallelizing the Borg MOEA. First, we derive theoretical models for predicting speedup, efficiency, and processor count lower and upper bounds. Second, we validate these models on a simple problem, DTLZ2, and a harder, non-separable problem, UF11. Third, we examine the effects of scaling on convergence speed and solution quality. These experiments are performed on the 62, 976 core Texas Advanced Computing Center (TACC) Ranger system.

A multi-objective decision-making approach to the journal submission problem

When researchers complete a manuscript, they need to choose a journal to which they will submit the study. This decision requires to navigate trade-offs between multiple objectives. One objective is to share the new knowledge as widely as possible. Citation counts can serve as a proxy to quantify this objective. A second objective is to minimize the time commitment put into sharing the research, which may be estimated by the total time from initial submission to final decision. A third objective is to minimize the number of rejections and resubmissions. Thus, researchers often consider the trade-offs between the objectives of (i) maximizing citations, (ii) minimizing time-to-decision, and (iii) minimizing the number of resubmissions. To complicate matters further, this is a decision with multiple, potentially conflicting, decision-maker rationalities. Co-authors might have different preferences, for example about publishing fast versus maximizing citations. These diverging preferences can lead to conflicting trade-offs between objectives. Here, we apply a multi-objective decision analytical framework to identify the Pareto-front between these objectives and determine the set of journal submission pathways that balance these objectives for three stages of a researcher’s career. We find multiple strategies that researchers might pursue, depending on how they value minimizing risk and effort relative to maximizing citations. The sequences that maximize expected citations within each strategy are generally similar, regardless of time horizon. We find that the “conditional impact factor”—impact factor times acceptance rate—is a suitable heuristic method for ranking journals, to strike a balance between minimizing effort objectives and maximizing citation count. Finally, we examine potential co-author tension resulting from differing rationalities by mapping out each researcher’s preferred Pareto front and identifying compromise submission strategies. The explicit representation of trade-offs, especially when multiple decision-makers (co-authors) have different preferences, facilitates negotiations and can support the decision process.

Auto-Adaptive Search Capabilities of the New Borg MOEA: A Detailed Comparison on Alternative Product Family Design Problem Formulations

There is a growing recognition that the design and management of complex engineered systems requires a fundamental advance in our ability to identify and exploit their inherent tradeoffs across a growing number of decisions and objectives. In support of this challenge, this study provides a rigorous evaluation of modern “many-objective” evolutionary optimization algorithms. The computational power of modern high-performance computing environments makes it possible to investigate optimization algorithm performance in ways that were not historically feasible. This study uses millions of algorithm runs, each performing hundreds of thousands of function evaluations, to do a Sobol’ global sensitivity analysis on algorithm parameterization. We present this analysis for two algorithms across four formulations of a General Aviation Aircraft (GAA) conceptual product family design problem. The two algorithms are the recently introduced Borg Multi-Objective Evolutionary Algorithm (MOEA), a promising auto-adaptive multi-operator search algorithm, and the e-MOEA, its algorithmic forebear. The four formulations of the GAA problem vary in their complexity and allow us to investigate the assumption that complex problem formulations are more difficult to solve.