Parallelized Direct Search of a Binary Objective

Many scientific and engineering problems are solved by utilizing simulations of computationally intensive mathematical models within massive design spaces. As a result, parametric studies of these models are cost prohibitive in terms of computational time. The focus of this work is a particular kind of parameter sweep problem where the simulation of the model, given real parameters, results in a binary value. The goal of this work is to design, develop, and implement a parallel algorithm for bounding a binary objective when the simulation is a computationally intensive mathematical model. A fully functioning implementation is provided for a two-dimensional example using client-server architecture. Results show that a straight bisection search is approximately 50% faster than a full parametric sweep for most continuous functions. With parallelization and load balancing, the simulation is remarkably faster, exhibiting near-linear speedup up to 16 processors for most functions. [DOI: 10.1115/1.3510589]