Preference-based expensive multi-objective optimization without using an ideal point
Abstract As a decision maker is not always interested in the entire Pareto front, a natural idea is to take into account user preferences in computationally expensive multi-objective optimization to focus on searching the preferred region. However, most existing methods rely on the estimated ideal p...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-06-01
|
| Series: | Complex & Intelligent Systems |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s40747-025-01905-w |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract As a decision maker is not always interested in the entire Pareto front, a natural idea is to take into account user preferences in computationally expensive multi-objective optimization to focus on searching the preferred region. However, most existing methods rely on the estimated ideal point. Since the search process of preference-based approaches is always biased towards the region of interest (ROI), it is highly challenging to estimate the ideal point in preference-based expensive multi-objective optimization. Incorrect estimation of the ideal point may lead to the search direction being far away from the ROI and also result in normalization errors for each objective. This paper proposes a preference-based surrogate-assisted algorithm, PMEGO, to overcome this issue. In the preference management module, a set of new reference points derived from the user reference point is used to construct a series of subproblems, with the key merit of not using the ideal point. The Gaussian process model is built on the objective functions. In the model-based optimization, the projection distance with upper confidence bound (UCB) is developed as the fitness of solutions for each subproblem. Finally, the expected achievement scalarizing function improvement (EASFI) is employed to further screen out the best solutions for evaluation using the real objective functions. Comparative experiments are conducted on ZDT, DTLZ, SDTLZ, as well as two real-world applications. The experimental results show that the proposed method is competitive compared to three state-of-the-art algorithms. |
|---|---|
| ISSN: | 2199-4536 2198-6053 |