Title: Developing multi-agent systems as a new approach for engineering design and optimization Speaker: John Siirola, Interview Candidate, Department of Chemical Engineering, Carnegie Mellon University Date/Time: Tuesday June 21, 2005, 10:00-11:00 am Location: Building 980, Room 95 (Sandia NM) Brief Abstract: The explosive growth of computing capacity raises a new question for the process systems engineering community: How can we take what is effectively an unlimited computing resource and apply it to solving new, larger, harder engineering problems? In this seminar, I present the framework and implementation of an agent-based system for general single- and multi-objective optimization. This system creates and manages a collaborative environment wherein a diverse library of many independently-running algorithms communicate through the sharing of intermediate and final solutions to an optimization problem. These algorithms span the entire range of available approaches, including rigorous gradient-based or branch-and-bound optimization methods, stochastic search methods, and heuristic methods. The agent system also introduces the new concept of "polymorphic optimization." Analogous to diversity in the algorithmic library, polymorphic optimization provides diversity in the formulation of the original problem. This formalism allows algorithms working on fundamentally different formulations of the original optimization problem to collaborate and seamlessly share solution information. My approach represents the confluence of two primary areas of optimization research: development of optimization algorithms and development of formulation methodologies. The former aims to create new, more robust algorithms that can solve a larger variety of problem classes and/or solve specific problem classes faster. The latter aims to develop new approaches for representing problems in a form that existing optimization algorithms will more likely solve successfully. Current research in both areas focuses primarily on "traditional" serial optimization: optimizing a single objective using one algorithm running on a stand-alone personal computer. In contrast, the emerging standard for high-performance computing is the distributed cluster of inexpensive personal computers. Historically, researchers have used clusters for large-scale parallel algorithms. Engineering optimization is a member of a sizable body of problems where, although there are several approaches to solving the problem, these approaches are primarily serial algorithms that do not readily lend themselves to classical parallelism. Agent-based systems form a promising approach to this next generation of parallel systems: parallelism at the method and formulation level. In this talk, I illustrate the benefits, future potential, and remaining challenges of an agent-based optimization approach through several examples, including the single-objective [1] and multi-objective [2] optimization of non-convex functions, and the design and multi-objective optimization of microfluidic separation systems [3,4]. References: [1] J.D. Siirola, S. Hauan, and A.W. Westerberg. "Toward Agents-Based Process Systems Engineering: Proposed Framework and Application to Non-convex Optimization." Comp.Chem.Engng. 27(12), pp. 1801-1811 (2003). |