Multi-Level Decomposition For Tractability In Structural Design Optimization

This paper describes two approaches that allow decomposition of structural design problems that enable structural design optimization to be performed on systems that are often seen as too large or complex to address in a single optimization. The COMPOSE method is shown to enable optimization of many tightly coupled subsystems despite the usual problem of non-convergence when a subsystem is repeatedly removed and optimized under the previously prevailing boundary conditions, then reinserted into the entire system model, which subsequently causes the boundary conditions experienced by the subsystem to change. Use of COMPOSE may allow reducing the number of whole-system evaluations necessary for component design from hundreds or thousands to fewer than ten, while still exploring the component design space extensively. A second approach, collaborative independent agents, is shown to address problems that have both large design spaces and time-intensive analyses, rendering them intractable to traditional methods. In an example problem, a set of loosely coupled optimization agents is shown to reduce dramatically the computing time needed to find good solutions to such problems. The savings result from the continuing transfer of results from rapid, low-refinement, less accurate search to agents that search at the full level of refinement and accuracy demanded of a solution of the problem, essentially providing them guidance as to what portions of the design space are likely to be worth searching in greater detail.