💻 computer science

Optimization is Not Enough: Why Problem Formulation Deserves Equal Attention

This paper demonstrates through a case study on laminated composite topology optimization that incorporating domain knowledge via a sequential problem formulation yields superior and more interpretable results compared to context-agnostic concurrent approaches, arguing that problem formulation deserves equal attention to algorithm optimization in black-box engineering design.

Original authors: Iván Olarte Rodríguez, Gokhan Serhat, Mariusz Bujny, Fabian Duddeck, Thomas Bäck, Elena Raponi

Published 2026-02-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Iván Olarte Rodríguez, Gokhan Serhat, Mariusz Bujny, Fabian Duddeck, Thomas Bäck, Elena Raponi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: It's Not Just About the Engine, It's About the Map

Imagine you are trying to find the fastest route across a country. You have a very powerful, high-tech GPS (the optimization algorithm) that can calculate millions of routes per second. However, if you give the GPS a map that is drawn upside down, or if you tell it to drive through a lake because you didn't specify that water is an obstacle, that super-powerful GPS will still fail. It will find the "fastest" way to drive through the lake, which is useless.

This paper argues that in engineering design, how you define the problem (the map) is just as important as the algorithm (the GPS) you use to solve it.

The authors are studying how to design strong, lightweight structures made of composite materials (like carbon fiber). These materials are special because they are strong in one direction but weak in another, kind of like a piece of wood that is easy to split along the grain but hard to break across it.

The Challenge: The "Black Box" Dilemma

In many engineering problems, we can't see exactly how the math works inside the computer simulation. It's a "black box." You put a design in, and it spits out a score (how strong or weak it is). Because we can't see the inside, we usually just throw our best algorithms at the problem and hope they find the answer.

The authors wanted to test two different ways of asking the computer to solve the design problem for a cantilever beam (a structure fixed at one end, like a diving board, holding a weight at the other).

The Two Strategies: "Do Everything at Once" vs. "Step-by-Step"

The team compared two approaches to designing this beam:

1. The Concurrent Approach (The "Do Everything at Once" Method)

How it works: You tell the computer, "Here are all the variables: where the material goes, how thick it is, and which way the fibers point. Figure it all out at the same time."
The Analogy: Imagine asking a chef to cook a complex 5-course meal while simultaneously building the dining table, painting the walls, and arranging the flowers. They have to juggle everything at once.
The Result: The computer gets overwhelmed. It often creates designs that look weird or physically impossible (like a bridge that isn't connected to the ground). It struggles to find the best balance because it's trying to solve too many puzzles simultaneously.

2. The Sequential Approach (The "Step-by-Step" Method)

How it works: You break the problem down.
- Step 1: "First, just figure out the best shape and where the material should go. Ignore the fiber directions for now; just assume they are average."
- Step 2: "Now that we have a solid shape, let's fine-tune the direction of the fibers to make it as strong as possible."
The Analogy: This is like the chef first building the table and setting the room (Step 1), and then cooking the meal (Step 2). Or, like an architect drawing the blueprints for a house before the electrician runs the wires.
The Result: This approach consistently produced better, stronger, and more realistic designs.

What They Found

The researchers tested this using several different "smart" computer algorithms (like CMA-ES, TuRBO, and others). Here is what happened:

The "Do Everything" method often led to suboptimal designs. The computer would get stuck in a "local trap," finding a solution that looked okay but wasn't the best, or it would create designs that wouldn't actually work in the real world (like a beam that falls apart).
The "Step-by-Step" method consistently found better solutions. By separating the "shape" problem from the "fiber direction" problem, the algorithms could focus their energy on one thing at a time.
The "Physics" Factor: The step-by-step method worked better because it used domain knowledge (real-world engineering rules). It respected the fact that the shape of the structure matters more than the fiber direction in the early stages.

The Takeaway

The paper concludes that optimization is not enough. Just having a faster or smarter algorithm doesn't matter if you ask the wrong question.

Don't treat the problem as a pure mystery: Even if you can't see the math inside the "black box," you should use your knowledge of physics to break the problem into logical steps.
Context matters: In engineering, variables aren't just random numbers; they represent real things (like wood grain or steel beams). Treating them as a single, messy pile of data leads to bad results. Treating them in the order nature intended (Shape first, then Material) leads to success.

In short: Before you hire the best detective (the algorithm), make sure you've organized the crime scene (the problem formulation) correctly.

Technical Summary: Optimization is Not Enough: Why Problem Formulation Deserves Equal Attention

1. Problem Statement

The paper addresses a critical gap in the field of black-box optimization (BBO): the tendency to focus on algorithmic performance in context-free setups while neglecting the impact of problem formulation and domain knowledge. In engineering design, particularly for laminated composite structures, optimization problems often involve expensive, non-differentiable simulations where gradients are unavailable.

The authors argue that optimization algorithms, regardless of sophistication, cannot inherently yield valid or practically useful solutions if the problem is not carefully formulated with appropriate physical modeling. Specifically, the study investigates the topology and fiber path optimization of laminated composite structures. The goal is to minimize structural compliance (maximize stiffness) for a cantilever beam under a volume constraint, while simultaneously optimizing the structural topology and the spatial distribution of fiber orientations.

2. Methodology

2.1 Problem Formulation and Parameterization

The study introduces a benchmark problem based on a linearly elastic cantilever beam subjected to a concentrated tip load. The design space is parameterized using two distinct methods:

Topology (MMC): The structural layout is defined using the Moving Morphable Components (MMC) method. This approach uses geometric parameters (center coordinates, length, thickness, and orientation) of deformable beam-like components to construct the structure, offering a compact and explicit geometric representation.
Material Stiffness (LPs): The anisotropic properties of the composite are modeled using Lamination Parameters (LPs). Specifically, the Lamination Parameter Interpolation Method (LPIM) is employed to enable continuous, spatially variable fiber orientations. This method ensures that the resulting stiffness variations are physically feasible and manufacturable by interpolating between design points on Miki's diagram.

The total design vector consists of 18 variables: 15 variables defining the topology (3 MMCs with symmetry assumptions) and 3 variables defining the LPs (volumetric ratio and two $V_3$ values at master nodes).

2.2 Optimization Strategies

To assess the impact of problem formulation, the authors compare two distinct strategies under a fixed function evaluation budget ( $B=1000$ ):

Concurrent Optimization: All 18 design variables (topology and material) are optimized simultaneously in a single 18-dimensional search space. This approach treats the problem as a pure black-box without leveraging physical insights regarding the separation of variables.
Sequential Optimization: The problem is decomposed into two stages:
- Stage 1: Optimize the 15 topology variables while fixing the material to a quasi-isotropic state. The budget is allocated proportionally to the dimensionality of the topology variables.
- Stage 2: Fix the optimal topology from Stage 1 and optimize the 3 LP variables using the remaining budget. This approach leverages domain knowledge by decoupling the structural layout from the material anisotropy.

2.3 Algorithms and Evaluation

Seven black-box optimization algorithms were tested under both strategies:

Population-based: CMA-ES (active variant), Differential Evolution (DE).
Surrogate-based (Bayesian Optimization): Vanilla BO, TuRBO-1, TuRBO-m, HEBO, and BAxUS.

The objective function is modified compliance, penalized for volume constraint violations and structural disconnection (using a connectivity penalty function $\Psi(x)$ ) to handle infeasible designs without crashing the solver.

3. Key Results

3.1 Performance Comparison

Sequential vs. Concurrent: The sequential strategy consistently yielded better-performing solutions (lower modified compliance) across most algorithms compared to the concurrent strategy. Statistical analysis (Wilcoxon signed-rank test) confirmed that the medians of the sequential approach were significantly better for TuRBO-1, TuRBO-m, and vanilla BO.
Algorithm Sensitivity: While the concurrent strategy showed faster initial convergence (due to the ability to adjust fiber orientations early), it often converged to suboptimal or non-physical designs. The sequential approach allowed algorithms to focus their search capacity effectively on the LP variables once the topology was established, leading to more robust exploitation of the convex LP sub-problem.
Best Performers: TuRBO-1 and TuRBO-m in the sequential setting achieved the best average performance for larger budgets. HEBO performed well in the early stages (first 400 evaluations) but saw its relative performance decline as the budget increased.

3.2 Qualitative Analysis of Solutions

Physical Realism: Concurrent strategies often produced designs with "excess material" in low-stress regions and inconsistent fiber orientations that did not align with principal stress directions. In contrast, sequential strategies produced more interpretable and physically coherent fiber paths (e.g., fibers aligning vertically near the clamp and shifting to $\pm 60^\circ$ near the load).
Search Behavior: In the concurrent formulation, acquisition functions tended to prioritize changes to topology variables (which have a stronger influence on the objective landscape), effectively biasing the search away from optimal LP configurations. The sequential approach eliminated this imbalance, allowing for targeted exploration of the material design space.

4. Significance and Claims

The paper posits that optimization is not enough; the formulation of the problem is equally critical. The authors claim that:

Domain Knowledge is Essential: Incorporating physical insights (such as the distinct nature of topology vs. material variables) into the optimization pipeline can significantly outperform "context-agnostic" black-box approaches.
Sequential Decomposition: For complex engineering problems with structured variables, a sequential approach that decouples variables based on their physical nature can lead to superior, more interpretable, and physically valid designs compared to simultaneous optimization.
Benchmarking Limitations: Current black-box benchmarks (like BBOB) often fail to reward physically informed strategies. The authors argue for the development of new benchmarks that evaluate not just algorithmic efficiency, but also the ability of strategies to leverage domain knowledge and produce realistic engineering solutions.

The study concludes that successful engineering optimization depends on a synergy between algorithm selection and problem formulation, urging the community to move beyond treating variables as arbitrary inputs in a purely black-box landscape.