PPSL-MOBO

Code for the AAAI 2026 paper: Parametric Pareto Set Learning for Expensive Multi-Objective Optimization.

Repository Overview

• mobo/: Contains modules for Gaussian Process surrogate models and their training, building upon the PSL-MOBO codebase.
• problems/: A collection of synthetic test problems used in our experiments. This includes:
  • Multi-objective problems with shared components.
  • Dynamic multi-objective optimization problems (DMOPs).
• results/: Directory containing the raw data from our experiments. Note: This will be removed after the review process.

Core Implementation

• model.py: Defines the neural network architecture for the Parametric Pareto Set Learner.
• trainer.py: Handles the training loop for the PPSL model under different parameterization strategies.
• baselines_mobo.py: A unified implementation of competitive MOBO algorithms used as benchmarks.
• experiment_mop_sc.py: Runs the primary experiments for problems with shared components.
• experiment_dmop.py: Runs the primary experiments for dynamic problems.
• utils.py: Provides helper functions, notably for computing gradients of the smooth Tchebycheff scalarization.

Cite

If you use this code, please cite the paper:

@inproceedings{cheng2026parametric,
  title={Parametric Pareto Set Learning for Expensive Multi-Objective Optimization},
  author={Cheng, Ji and Xue, Bo and Zhang, Qingfu},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={40},
  number={43},
  pages={36829--36837},
  year={2026}
}