Asynchronous Parallel Pattern Search for Nonlinear Optimization

Abstract

We introduce a new asynchronous parallel pattern search (APPS). Parallel pattern search can be quite useful for engineering optimization problems characterized by a small number of variables (say, fifty or less) and by objective functions that are expensive to evaluate, such as those defined by complex simulations that can take anywhere from a few seconds to many hours to run. The target platforms for APPS are the loosely coupled parallel systems now widely available. We exploit the algorithmic characteristics of pattern search to design variants that dynamically initiate actions solely in response to messages, rather than routinely cycling through a fixed set of steps. This gives a versatile concurrent strategy that allows us to effectively balance the computational load across all available processors. Further, it allows us to incorporate a high degree of fault tolerance with almost no additional overhead. We demonstrate the effectiveness of a preliminary implementation of APPS on both standard test problems as well as some engineering optimization problems.