[Paper Review] Dealing with Structure Constraints in Evolutionary Pareto Set Learning

Xi Lin, Xiaoyuan Zhang+2 more|arXiv (Cornell University)|Oct 31, 2023

Advanced Multi-Objective Optimization Algorithms4 citations

TL;DR

This paper proposes Evolutionary Pareto Set Learning (EPSL), a model-based evolutionary framework that learns the entire Pareto set for continuous multiobjective optimization problems while incorporating user-defined structure constraints. By generalizing MOEA/D using a neural network to represent the Pareto set and optimizing it via evolutionary stochastic gradient descent, EPSL enables decision-makers to trade off Pareto optimality with desired solution set structures—demonstrating superior performance and structural simplicity on benchmark and real-world problems.

ABSTRACT

In the past few decades, many multiobjective evolutionary optimization algorithms (MOEAs) have been proposed to find a finite set of approximate Pareto solutions for a given problem in a single run, each with its own structure. However, in many real-world applications, it could be desirable to have structure constraints on the entire optimal solution set, which define the patterns shared among all solutions. The current population-based MOEAs cannot properly handle such requirements. In this work, we make the first attempt to incorporate the structure constraints into the whole solution set by a single Pareto set model, which can be efficiently learned by a simple evolutionary stochastic optimization method. With our proposed method, the decision-makers can flexibly trade off the Pareto optimality with preferred structures among all solutions, which is not supported by previous MOEAs. A set of experiments on benchmark test suites and real-world application problems fully demonstrates the efficiency of our proposed method.

Motivation & Objective

• To address the limitation of finite population-based MOEAs in approximating the full Pareto set, especially under structure constraints.
• To develop a model-based evolutionary method that learns the entire Pareto set as a continuous, parameterized manifold rather than a finite set of solutions.
• To enable decision-makers to actively incorporate preferred structural patterns (e.g., shared components, polygonal chains) into the solution set during optimization.
• To generalize decomposition and collaboration in MOEA/D using a single differentiable model and evolutionary optimization.
• To demonstrate that structured Pareto sets can achieve nearly comparable hypervolume and IGD+ performance with significantly simpler, more interpretable structures.

Proposed method

• EPSL uses a neural network model $ h_{\boldsymbol{\theta}}(\boldsymbol{\lambda}) $ parameterized by $ \boldsymbol{\theta} $ to map preference vectors $ \boldsymbol{\lambda} \in \mathbb{R}^m $ to decision variables $ \boldsymbol{x} \in \mathbb{R}^n $, forming a continuous Pareto set approximation.
• The method generalizes MOEA/D’s scalarization by replacing fixed subproblems with a single continuous model that handles all preferences simultaneously.
• An evolutionary stochastic gradient descent method is employed to optimize the model parameters $ \boldsymbol{\theta} $, enabling joint learning across all preferences.
• Structure constraints—such as polygonal chains or shared components—are embedded into the model via regularization or architectural constraints on the solution set.
• The framework uses a preference-conditioned forward pass: $ \boldsymbol{\lambda} \rightarrow h_{\boldsymbol{\theta}}(\boldsymbol{\lambda}) \rightarrow \boldsymbol{x} \rightarrow \boldsymbol{F}(\boldsymbol{x}) $, with gradients backpropagated through the model.
• The optimization is performed using Monte Carlo sampling over preference vectors, allowing efficient gradient-based training with the same computational budget as standard MOEAs.

Experimental results

Research questions

• RQ1 Can a model-based evolutionary framework learn the entire Pareto set for black-box multiobjective optimization problems more effectively than finite population MOEAs?
• RQ2 How can structure constraints—such as shared components or polygonal chain patterns—be effectively embedded into the Pareto set during optimization?
• RQ3 To what extent can a simplified, structured Pareto set maintain competitive convergence and convergence diversity compared to the ground truth Pareto set?
• RQ4 Can the proposed method achieve a favorable trade-off between Pareto optimality and solution set interpretability through user-defined structural priors?
• RQ5 How does the performance of EPSL with structure constraints compare to unconstrained EPSL and other state-of-the-art MOEAs on real-world engineering problems?

Key findings

• EPSL with polygonal chain structure constraints achieved a median hypervolume gap (ΔHV) of 8.23e-03 on the four-bar truss design problem (RE21), compared to 2.26e-03 for unconstrained EPSL, indicating a small performance trade-off for significant structural simplification.
• On the RE25 problem, EPSL with polygonal chain constraints achieved a ΔHV of 4.29e-08 and IGD+ of 1.24e-07, showing near-identical performance to the unconstrained version.
• For the synthetic problem (38), EPSL successfully learned a polygonal chain structure that closely approximated the ground truth Pareto set, with visually similar Pareto fronts.
• The constrained Pareto sets had much simpler structures than the unconstrained ones, yet their corresponding Pareto fronts provided nearly comparable trade-offs, validating the effectiveness of the structure constraint trade-off.
• In all tested engineering problems, EPSL with structure constraints maintained competitive convergence and diversity metrics while producing more interpretable and explainable solution sets.
• The results confirm that EPSL can effectively learn a Pareto set with desired structural patterns while maintaining high-quality approximation to the true Pareto front.