Improving the Robustness of the Projected Gradient Descent Method for Nonlinear Constrained Optimization Problems in Topology Optimization

Lucka Barbeau, Marc-Étienne Lamarche-Gagnon, Florin Ilinca, Barbeau L, Lamarche-Gagnon M.-É, Ilinca F · 2024

Read Improving the Robustness of the Projected Gradient Descent Method for Nonlinear Constrained Optimization Problems in Topology Optimization by Lucka Barbeau, Marc-Étienne Lamarche-Gagnon, Florin Ilinca, Barbeau L, Lamarche-Gagnon M.-É, Ilinca F (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-07634.txt.

Introduction

Lucka Barbeau, Marc-Étienne Lamarche-Gagnon, and Florin Ilinca harden a gradient method for design optimization in a 2024 paper. The method, projected gradient descent, is simple and scales to huge design spaces, which suits topology work. Nonlinear limits, though, can shake its steps.

Their first fix folds simple bound limits into the projection step with a cleaner active set routine. The second splits the update along the constraint directions and adjusts after projection using a Lagrange style estimate. On a heat sink design task, the tuned method matched or beat the standard rival solver. It did so with little parameter fuss, the authors note.

Catalog record

CREATORLucka Barbeau, Marc-Étienne Lamarche-Gagnon, Florin Ilinca, Barbeau L, Lamarche-Gagnon M.-É, Ilinca F
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE88 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-07634

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Improving the Robustness of the Projected Gradient Descent Method for Nonlinear Constrained Optimization Problems in Topology Optimization Lucka Barbeau, Marc-Étienne Lamarche-Gagnon, Florin Ilinca, Barbeau L, Lamarche-Gagnon M.-É, Ilinca F Abstract The Projected Gradient Descent (PGD) algorithm is a widely used and efficient first-order method for solving constrained optimization problems due to its simplicity and scalability in large design spaces. Building on recent advancements in the GD algorithm—where an inertial step component has been introduced to improve efficiency in solving constrained optimization problems—this study introduces two key enhancements to further improve the algorithm’s performance and adaptability in large-scale design spaces. First, univariate constraints (such as design variable bounds constraints) are directly incorporated into the projection step via the Schur complement and an improved active set algorithm with bulk constraints manipulation, avoiding issues with min–max clipping. Second, the update step is decomposed relative to the constraint vector space, enabling a post-projection adjustment based on the state of the constraints and an approximation of the Lagrangian, significantly improving the algorithm’s robustness for problems with nonlinear constraints. Applied to a topology optimization problem for heat sink design, the proposed PGD algorithm demonstrates performance comparable to or exceeding that of the Method of Moving Asymptotes (MMA), with minimal parameter tuning. These results position the enhanced PGD as a robust tool for complex optimization problems with large variable space, such as topology optimization problems. Publication type: Article Type Journal: International Journal for Numerical Methods in Engineering Volume: 00 keywords Inertial projected gradient descent, Infeasible active set, Bulk constraints manipulation, Schur complement decomposition, Step decomposition, Nonlinear constraints, Topology optimization † † copyright-year: 2023 † † startpage: 1 † † titlemark: Improving the Robustness of the Projected Gradient Descent Method for Nonlinear Constrained Optimization Problems in Topology Optimization † † corresponding: Corresponding author Lucka Barbeau, 75 Bd de Mortagne, Boucherville, QC, J4B 6Y4, Canada † † citation: . † † copyright: , , and . † † ctitle: Improving the Robustness of the Projected Gradient Descent Method for Nonlinear Constrained Optimization Problems in Topology Optimization. † † cjournal: International Journal for Numerical Methods in Engineering. † † cvolume: 2024;00(00):1–18 Barbeau et al. \presentaddress This is sample for present address text this is sample for present address text. 1 Introduction Topology optimization is a mathematical tool that determines the optimal structure or material distribution for a given task. It has seen a significant increase in interest due to advancements in manufacturing technologies, such as 3D printing, which can produce the complex geometries that often result from this process [ 1 ] . Applications of topology optimization algorithms span a diverse range of fields, including structural engineering [ 2 ] , heat transfer [ 3 ] , and fluid dynamics [ 4 ] , among others. At the core of the topology optimization algorithm is the optimizer that minimizes a given cost function by iteratively changing the topology while preserving certain constraints related to the geometry (volume, perimeter projection, etc.) or some physical properties of the topology (compliance, pressure drop, etc.). A large variety of algorithms exist to solve such constrained optimization problems. The most known and widely spread in the field of topology optimization are the Method of Moving Asymptotes (MMA) [ 5 ] , the Sequential Linear Programming (SLP) [ 6 ] , the Sequential Quadratic Programming (SQP) [ 7 ] , the optimality criteria [ 8 ] , the Interior Point Method (IPM) [ 9 ] , and the Projected Gradient Descent (PGD) [ 10 , 11 ] . Among these, projected gradient descent has seen a recent increase in interest, along with other first-order methods, due to its simplicity, reduced cost, and better scaling compared to second-order methods [ 11 ] . The PGD method benefits from having a generally small set of tunable parameters, making it easy to employ. For example, the work of Chzhen et al. modified and improved this algorithm to make it parameter-free [ 12 ] . Furthermore, work was done to improve the efficiency of this algorithm for topology optimization problems by using an acceleration method based on the inertial principal [ 11 , 13 ] . However, considering multiple constraints directly in the projection step leads to a general quadratic programming (QP) problem requiring appropriate methods to solve ( i.e. , augmented Lagrange multiplier, active set method, interior point method). However, the projection step in the PGD algorithm involves solving a general convex QP problem, often a computationally intensive task. To manage this complexity, an active set approach is employed, which reduces the size of the projection matrix by focusing only on the active constraints [ 14 ] . The active set methodology was effectively applied in a PGD context by Dostal et al. [ 15 ] , and since then, substantial advances have refined how active sets are incorporated within projections. One of the key developments is the infeasible strategy, where the active set of constraints is dynamically adjusted by adding constraints that are violated and removing those with negative Lagrange multipliers. This technique allows the algorithm to proceed from an initially infeasible point, gradually refining the feasible region until it reaches optimality [ 16 , 17 , 18 ] . To accelerate the convergence, some algorithms introduce methods to deal with multiple constraints simultaneously. Schittkowski [ 19 ] proposes an approach that combines bulk constraint manipulation with a step length adjustment, removing redundant constraints based on activity duration to avoid cycling. Cristofari et al. [ 20 ] proposed a two-stage algorithm for univariate constraints (bounds constraints) that enables backtracking and readjustment of the active set if updates fail to improve the objective, followed by a line search to ensure progress toward the solution. Additionally, to accelerate the definition of the active set, stochastic algorithms have also seen some development in this approach, manipulating multiple constraints simultaneously while ensuring, due to the stochastic nature of the algorithm, the convergence of the method [ 14 ] . However, the robustness of the PGD method, especially in the presence of nonlinear constraints, can still be improved, which would improve the applicability of this algorithm to complex topology optimization problems. The use of nonlinear constraints in topology optimization problems is becoming increasingly frequent. As an example, the addition of overhang constraints [ 21 , 22 , 23 ] , which aims at enhancing the manufacturability of the design of the part through topology optimization, introduces nonlinear constraints that can affect the convergence of the optimizer [ 24 ] . Some algorithm modifications are required to handle these constraints in the PGD method. The active set approach generally also requires either a stochastic algorithm to manipulate multiple constraints simultaneously or a sequence sequential manipulation of constraints, which can drastically increase the cost. Additionally, there is room for improvement in handling univariate constraints to enhance the efficiency of the PGD algorithm. Univariate constraints (usually bounds constraints), which apply to only one of the design variables, are ubiquitous in topology optimization problems due to the requirement of maintaining the phase indicator within the bound of the material model used (usually between 0 and 1 for density-based approaches). As the discretization of the physics of the topology optimization problem at hand leads to a large number

Read the complete text