A Note on the Convergence of Analytical Target Cascading With Infinite Norms Jeongwoo Han Argonne National Laboratory, 9700 S. Cass Avenue, Building 362, Argonne, IL 60439 e-mail: jhan@anl.gov Panos Y. Papalambros Department of Mechanical Engineering, University of Michigan, 2250 GG Brown Building, Ann Arbor, MI 48104 e-mail: pyp@umich.edu Analytical target cascading (ATC) is a multidisciplinary design optimization method for multilevel hierarchical systems. To im- prove computational efficiency, especially for problems under un- certainty or with strong monotonicity, a sequential linear pro- gramming (SLP) algorithm was previously employed as an alternate coordination strategy to solve ATC and probabilistic ATC problems. The SLP implementation utilizes L norms to maintain the linearity of SLP subsequences. This note offers a proof that there exists a set of weights such that the ATC algo- rithm converges when L norms are used. Examples are also pro- vided to illustrate the effectiveness of using L norms as a penalty function to maintain the formulation linear and differentiable. The examples show that the proposed method provides more robust results for linearized ATC problems due to the robustness of the linear programming solver. DOI: 10.1115/1.4001001 1 Introduction Design optimization of complex systems often requires multi- disciplinary analyses involving significant interactions. Such problems may be solved with an all-in-one AIOmethod in which the system is treated as a fully integrated single problem. When the analysis models used at each optimization iteration are computationally expensive or difficult to solve, the AIO method may not be practical or reliable. In such cases, decomposition strategies can be used: the problem is partitioned into several sub- problems that are solved with an appropriate coordination strat- egy. Decomposition strategies are classified as nonhierarchical or hierarchical Figs. 1aand 1b. Nonhierarchical partitions often use two levels: subproblems, typically representing different sub- systems or disciplinary analyses, are optimized concurrently, while a system-level problem coordinates the interactions among the subsystems 1–8. On the other hand, multilevel hierarchical systems are typically partitioned into physical subsystems or ob- jects. Hierarchical systems can be optimized using coordination strategies such as analytical target cascading ATCand its exten- sion, probabilistic analytical target cascading PATC9,10. In Figs. 1aand 1b, each block in the hierarchical structure is referred to as an element and is an optimization subproblem. Let quantities with indices ij be related to element j at level i. An element is coupled with its parent and children via targets t ij and responses r ij , and ATC allows relaxed consistency constraints t ij = r ij using penalty functions t ij - r ij . Then, a subproblem of element j at level i can be expressed as follows: min x ij f ij x ij + t ij - r ij subject to g ij x ij 0, h ij x ij = 0 where x ij = x ij , r ij , t i+1k , k C ij j E i , i = p,..., s 1 where f ij , g ij , and h ij are the separated objective, inequality, and equality constraints of element ij, respectively; C ij is the set of children of element ij. Using a penalty function that decreases to zero monotonically as the inconsistency becomes smaller, ATC enforces the consis- tency of values shared between elements. The proper choice of penalty functions and associated weights is critical for solution convergence. For quadratic penalty QPfunctions, large weights are required to obtain accurate and consistent solutions 11. Simi- lar to other decomposition strategies, ATC typically is more ex- pensive than the AIO method, if the latter could be used to obtain a solution, due to the coordination overhead. Michalek and Pa- palambros 12proposed an efficient weight update method WUMthat finds minimal weights to achieve a given level of consistency, especially important for problems with unattainable system targets. Still, the inner loop coordination, where the de- composedATC problems are solved iteratively, is computationally expensive. To address this, Tosserams et al. 13introduced a separable augmented Lagrangian ALpenalty function and an alternating direction solution method ALAD, resulting in signifi- cant computational cost savings. Also, an ATC dual coordination algorithm was proposed using the Lagrangian duality theory, where Lagrangian dual problems are solved by subgradient opti- mization to update Lagrange multipliers 14,15. In order to take advantage of parallel computing, diagonal quadratic approxima- tion DQAand truncated diagonal quadratic approximation TDQAwere applied by linearizing the cross term of the aug- mented Lagrangian function 16. To improve the computational efficiency further, especially for problems under uncertainty or with strong monotonicity, a se- quential linear programming SLPalgorithm by Fletcher et al. 17was previously adapted as an alternate coordination strategy to solve ATC and PATC problems 18–20. This coordination strategy was able to reduce the computational cost by taking ad- vantage of the strong monotonicity of problems and the simplicity and ease of uncertainty propagation for a linear system. The SLP- based ATC algorithm utilizes L norms to maintain the linearity of SLP subsequences. With L norms, a subproblem of element j at level i can be expressed as follows: min x ij , ij , i+1k f ij x ij + ij + kC ij i+1k subject to g ij x ij 0, h ij x ij = 0 - ij w t - r ij  ij Contributed by the Design Automation Committee of ASME for publication in the JOURNAL OF MECHANICAL DESIGN. Manuscript received January 14, 2009; final manu- script received December 18, 2009; published online March 1, 2010. Special Editor: Timothy W. Simpson. Fig. 1 aNonhierarchical and bhierarchical decompositions Journal of Mechanical Design MARCH 2010, Vol. 132 / 034502-1 Copyright © 2010 by ASME Downloaded 23 Mar 2010 to 141.212.126.174. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm