Abstract: In many real-world applications, the Pareto Set (PS) of a continuous multiobjective optimization problem can be a piecewise continuous manifold. A decision maker may want to find a solution set that approximates a small part of the PS and requires the solutions in this set share some similarities. This paper makes a first attempt to address this issue. We first develop a performance metric that considers both optimality and variable sharing. Then we design an algorithm for finding the model that minimizes the metric to meet the user's requirements. Experimental results illustrate that we can obtain a linear model that approximates the mapping from the preference vectors to solutions in a local area well.
Introduction
In Multiobjective Optimization Problems (MOPs), the goal is typically to find the Pareto Set (PS)—the collection of solutions where improving one objective necessitates degrading another. While mathematically elegant, the PS often forms a complex, continuous manifold. In practical engineering and manufacturing contexts, a decision-maker might find a "perfect" Pareto solution impractical. If every slight change in preference requires a completely different set of design parameters, manufacturing costs can skyrocket.
Engineers prefer variable sharing: solutions that are nearly optimal but maintain identical components or parameters across different product versions, supporting modular design and reusability. This paper addresses the gap by explicitly trading off a degree of Pareto optimality for enhanced variable sharing through a novel linear approximation approach.
Methodology: MOEA/D-LLA
The authors propose MOEA/D-LLA (Multi-Objective Evolutionary Algorithm based on Decomposition with Local Linear Approximation). This algorithm models the relationship between user preferences ($\lambda$) and solutions ($x$) within a local neighborhood.
The core of the approach is a linear model: $h_\theta(\lambda) = A(\lambda - \lambda^0) + b$. Here, $b$ represents a bias vector and $A$ is a matrix whose rows determine how decision variables change with preferences. To enforce variable sharing, the authors utilize (2,1)-norm regularization on matrix $A$. This is a classic machine learning technique to induce row-sparsity. If a row in $A$ is zeroed out, the corresponding decision variable remains constant (determined by the bias $b$) regardless of preference shifts, effectively "sharing" that variable across different solutions.
The performance metric $\mathcal{M}$ integrates optimality (measured via Chebyshev aggregation) and the degree of sharing (weighted by a penalty parameter $\gamma$). The algorithm combines standard genetic operators with predictions from this linear model to generate and refine candidate solutions.
Experimental Evaluation
The approach was evaluated on standard benchmarks (ZDT 1, 2, 4, 6 and DTLZ 1, 2, 3, 4) and a custom problem, MOZDT1. The MOZDT1 problem was specifically designed to have no natural variable sharing in its Pareto Set, serving as a "stress test" for the algorithm's ability to force modularity even when the underlying mathematics does not naturally support it.
Results and Discussion
Experimental results demonstrate that the linear model often improves convergence speed and accuracy on standard problems, likely because many benchmarks naturally exhibit some variable sharing. On 7 out of 8 standard instances, the LLA model outperformed the baseline MOEA/D-DE in terms of Inverted Generational Distance (R-IGD) and Hypervolume (R-HV).
The parameter $\gamma$ serves as a "modularity dial." As shown in the comparison below, increasing $\gamma$ forces the model to prioritize variable sharing at the cost of slight deviations from the true Pareto Front.
On the custom MOZDT1 problem, the algorithm successfully reduced decision variable variance as $\gamma$ increased. At $\gamma=5$, the predicted solutions became nearly identical, proving that the algorithm can effectively prioritize modularity even when it conflicts with pure Pareto optimality. Furthermore, the regression task used to train the linear model is convex and can be solved efficiently using regularized least squares.
Conclusion
This research provides a first attempt at approximating Pareto Set segments with a focus on variable sharing. While currently limited to local approximations around a specific preference vector, the use of a linear model with (2,1)-norm regularization offers a computationally efficient way to balance performance and modularity. Future work will explore global approximations and more complex, non-linear relationships using advanced learning techniques like deep reinforcement learning.
Summary updated