- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces Pareto Flow Learning, a Pareto set learning technique that formulates each scalarized sub-problem as a continuous-time Pareto gradient flow, learning to map a preference vector to a near-optimal solution. The authors propose to learn this mapping via meta-learning, where the flow is conditioned on a two-part task embedding. The first part of the task embedding is computed from the training set, while the second part includes a lower-dimensional target-task-specific embedding. During meta-training, the target-specific embedding is learned via bilevel optimization. Lastly, the authors propose to use the learned flow-mapping in Multi-Objective Bayesian Optimization (MOBO), where the pretrained flow-mapping is used to optimize the acquisition function, moving the optimization from the decision space to the conditioned solution manifold. Furthermore, the authors also show how the task-specific encoding can be fine-tuned during active learning.
Overall the paper is well motivated and tackles an interesting problem. The empirical evaluation show that the i) proposed approach improves both the convergence speed and asymptotic performance over existing MOBO methods, ii) the Pareto flow formulation provides a better learning signal than using only terminal loss, and iii) the task embedding module has significant impact on the final performance. However, the empirical evaluations are limited mostly synthetic problems, with low dimensional decision spaces, limiting the generalizability of the claims.
Critical strengths
- The paper introduces a Pareto gradient flow-based objective, which is extended to meta-learning setting and provide technical details on how it can be computed efficiently. The authors showcase empirically that the flow-based objective provides a more effective learning target than directly minimizing the terminal error.
- The authors divide the learned task embedding into pretrained task embedding and target task-specific lower-dimensional embedding . This allows one to finetune only a small part of the model parameters during active learning, improving the efficiency. Moreover, the authors recognize this as a crucial component for improving the performance of their proposed method.
Critical weaknesses
My main concerns are related to evaluations:
- The authors consider only synthetic functions and 4 engineering problems, all of which have low-dimensional decision spaces. One of the claims in the Bayesian Optimization section is that the acquisition function optimization is easier in the conditioned manifold. While this may be true, I would like to see the method evaluated in more complicated problems with higher number of decision variables, such as trajectory planning [1] or vehicle design problem [2] to further support this claim.
- The authors only consider hypervolume as their performance metric; While this metric is commonly used in the literature, it is known to be quite sensitive to the used reference point [3]. It would be good to consider other metrics such as IGD+ and R2 to support the claims, especially since the ground truth solution sets are available for the synthetic tasks. Moreover, the Pareto-Front visualizations show only the proposed method, making it impossible to determine how e.g. diverse the generated solutions are compared to the baselines.
[1] Wang, Z., Gehring, C., Kohli, P. & Jegelka, S.. Batched Large-scale Bayesian Optimization in High-dimensional Spaces. In Proceedings of the Twenty-First International Conference on Artificial Intelligence and Statistics, 2018
[2] T. Kohira, H. Kemmotsu, O. Akira, and T. Tatsukawa. Proposal of benchmark problem based on real-world car structure design optimization. In Proceedings of the Genetic and Evolutionary Computation Conference Companion, 2018
[3] H. Ishibuchi, R. Imada, Y. Setoguchi and Y. Nojima, How to Specify a Reference Point in Hypervolume Calculation for Fair Performance Comparison, in Evolutionary Computation, vol. 26, no. 3, pp. 411-440, 2018
Questions for authors
- The meta-loss function (Eq. 11) consists of many terms. How were the coefficients selected for the loss, and how sensitive is the loss is to these coefficients?
- The proposed approach seems to underperform in mDTLZ1 and mDTLZ$^{-1}$, where the Pareto Front is a simplex. Is there any intuition why this is the case?
- Table 2 recognizes the target task adaption as an important component of the mPFM-MOBO. Could you clarify if any of the baselines used a similar mechanism? If not, could a similar mechanism be applied to existing transfer PSL methods such as ExTrEMO or invTrEMO?
See also weaknesses section.
7 comments
Response (2/5)
Response to W2 (Part 1)
To provide a more comprehensive evaluation beyond hypervolume, we have additionally reported IGD+ and R2 on the synthetic benchmarks, covering in-distribution and out-of-distribution settings. We also have updated the results and the recommended reference in Appendix J of the updated manuscript.
(1) IGD+
In-distribution Settings
| Problem | mPFM-MOBO | DirHV-EGO | ExTrEMO | InvTrEMO | PSL-MOBO | PSL-DirHV-EGO |
|---|---|---|---|---|---|---|
| mDTLZ1 | 0.0297±0.0051 | 0.3470±0.1167 | 0.2019±0.0450 | 0.1873±0.1205 | 0.2522±0.0831 | 0.6020±0.1599 |
| mDTLZ2 | 0.0391±0.0111 | 0.1562±0.0136 | 0.1708±0.0165 | 0.1342±0.0469 | 0.1162±0.0241 | 0.2512±0.0346 |
| mDTLZ3 | 0.0411±0.0070 | 0.2424±0.0474 | 0.2172±0.0265 | 0.1857±0.0920 | 0.1824±0.0275 | 0.4158±0.0789 |
| mDTLZ4 | 0.0355±0.0028 | 0.1502±0.0144 | 0.1811±0.0203 | 0.1521±0.0339 | 0.1340±0.0200 | 0.2516±0.0381 |
| mDTLZ1⁻¹ | 0.0223±0.0013 | 0.2023±0.0354 | 0.1661±0.0265 | 0.0769±0.0390 | 0.1276±0.0408 | 0.3988±0.1697 |
| mDTLZ2⁻¹ | 0.0285±0.0018 | 0.0815±0.0061 | 0.0879±0.0068 | 0.0694±0.0135 | 0.0457±0.0051 | 0.2232±0.0426 |
| mDTLZ3⁻¹ | 0.0313±0.0030 | 0.1437±0.0664 | 0.1157±0.0112 | 0.0654±0.0075 | 0.0605±0.0123 | 0.2414±0.0466 |
| mDTLZ4⁻¹ | 0.0388±0.0041 | 0.0823±0.0111 | 0.1237±0.0161 | 0.0764±0.0187 | 0.0337±0.0017 | 0.2723±0.0449 |
Out-of-distribution Settings
| Problem | mPFM-MOBO | DirHV-EGO | ExTrEMO | InvTrEMO | PSL-MOBO | PSL-DirHV-EGO |
|---|---|---|---|---|---|---|
| mDTLZ1 | 0.0630±0.0140 | 0.7950±0.1146 | 0.4029±0.1280 | 0.7581±0.4914 | 0.7180±0.1336 | 0.9751±0.1491 |
| mDTLZ2 | 0.0436±0.0113 | 0.1331±0.0155 | 0.1959±0.0273 | 0.1923±0.0732 | 0.1150±0.0234 | 0.2403±0.0427 |
| mDTLZ3 | 0.0505±0.0116 | 0.3063±0.0499 | 0.2377±0.0267 | 0.2888±0.1294 | 0.2406±0.0431 | 0.4331±0.0490 |
| mDTLZ4 | 0.0371±0.0037 | 0.1402±0.0142 | 0.1981±0.0271 | 0.2721±0.0688 | 0.1202±0.0135 | 0.2448±0.0280 |
| mDTLZ1⁻¹ | 0.0960±0.1578 | 0.6818±0.1231 | 0.3562±0.0791 | 0.3288±0.3524 | 0.4560±0.1535 | 0.7276±0.1393 |
| mDTLZ2⁻¹ | 0.0331±0.0049 | 0.1006±0.0074 | 0.1099±0.0091 | 0.0999±0.0194 | 0.0578±0.0067 | 0.2167±0.0325 |
| mDTLZ3⁻¹ | 0.0323±0.0026 | 0.2084±0.0336 | 0.1413±0.0184 | 0.1109±0.0143 | 0.0958±0.0304 | 0.3164±0.0359 |
| mDTLZ4⁻¹ | 0.0346±0.0036 | 0.0781±0.0086 | 0.1235±0.0132 | 0.0914±0.0111 | 0.0365±0.0054 | 0.2481±0.0388 |
(2) R2
In-distribution Settings
| Problem | mPFM-MOBO | DirHV-EGO | ExTrEMO | InvTrEMO | PSL-MOBO | PSL-DirHV-EGO |
|---|---|---|---|---|---|---|
| mDTLZ1 | 0.0379±0.0012 | 0.0856±0.0127 | 0.0817±0.0102 | 0.0635±0.0197 | 0.0712±0.0119 | 0.1263±0.0214 |
| mDTLZ2 | 0.0966±0.0013 | 0.1125±0.0025 | 0.1135±0.0021 | 0.1073±0.0063 | 0.1159±0.0106 | 0.1343±0.0215 |
| mDTLZ3 | 0.0969±0.0010 | 0.1251±0.0068 | 0.1196±0.0042 | 0.1128±0.0110 | 0.1262±0.0119 | 0.1527±0.0272 |
| mDTLZ4 | 0.0953±0.0004 | 0.1101±0.0020 | 0.1132±0.0020 | 0.1090±0.0037 | 0.1078±0.0019 | 0.1214±0.0050 |
| mDTLZ1⁻¹ | 0.1043±0.0018 | 0.2128±0.0225 | 0.1977±0.0152 | 0.1286±0.0187 | 0.1694±0.0305 | 0.2749±0.0408 |
| mDTLZ2⁻¹ | 0.1250±0.0006 | 0.1339±0.0013 | 0.1451±0.0026 | 0.1371±0.0044 | 0.1286±0.0008 | 0.1976±0.0145 |
| mDTLZ3⁻¹ | 0.1257±0.0011 | 0.1611±0.0365 | 0.1529±0.0035 | 0.1354±0.0022 | 0.1334±0.0029 | 0.2051±0.0139 |
| mDTLZ4⁻¹ | 0.1299±0.0013 | 0.1364±0.0027 | 0.1548±0.0036 | 0.1391±0.0069 | 0.1281±0.0006 | 0.2049±0.0141 |
Response (3/5)
Response to W2 (Part 2)
Out-of-distribution Settings
| Problem | mPFM-MOBO | DirHV-EGO | ExTrEMO | InvTrEMO | PSL-MOBO | PSL-DirHV-EGO |
|---|---|---|---|---|---|---|
| mDTLZ1 | 0.0460±0.0032 | 0.1406±0.0115 | 0.1069±0.0114 | 0.1378±0.0651 | 0.1328±0.0185 | 0.1579±0.0115 |
| mDTLZ2 | 0.0981±0.0015 | 0.1090±0.0025 | 0.1187±0.0042 | 0.1127±0.0082 | 0.1211±0.0116 | 0.1397±0.0236 |
| mDTLZ3 | 0.0989±0.0016 | 0.1282±0.0054 | 0.1229±0.0056 | 0.1239±0.0156 | 0.1389±0.0129 | 0.1576±0.0259 |
| mDTLZ4 | 0.0955±0.0005 | 0.1094±0.0018 | 0.1160±0.0030 | 0.1221±0.0075 | 0.1060±0.0017 | 0.1216±0.0052 |
| mDTLZ1⁻¹ | 0.1319±0.0471 | 0.3107±0.0030 | 0.2853±0.0255 | 0.2040±0.0897 | 0.2916±0.0339 | 0.3121±0.0037 |
| mDTLZ2⁻¹ | 0.1263±0.0016 | 0.1417±0.0027 | 0.1531±0.0039 | 0.1472±0.0063 | 0.1334±0.0022 | 0.1930±0.0099 |
| mDTLZ3⁻¹ | 0.1257±0.0009 | 0.1844±0.0161 | 0.1645±0.0055 | 0.1501±0.0048 | 0.1467±0.0114 | 0.2295±0.0137 |
| mDTLZ4⁻¹ | 0.1285±0.0014 | 0.1350±0.0018 | 0.1554±0.0042 | 0.1446±0.0037 | 0.1285±0.0017 | 0.1969±0.0116 |
The additional results are largely consistent with our HV-based findings. In particular, mPFM-MOBO achieves the best IGD+ and R2 values on most synthetic tasks, demonstrating its effectiveness in approximating the Pareto front under limited evaluation budgets.
Regarding the Pareto front visualizations, we clarify that the original figures illustrate the performance of the learned Pareto set models. Among the baselines, PSL-MOBO, PSL-DirHV-EGO, and invTrEMO can generate such Pareto set model through their learning component, whereas the remaining methods do not provide an equivalent preference-conditioned solution generator. We have therefore added visual comparisons with these three methods in Appendix E, Figure 8 - 11, allowing a more direct assessment of the convergence and diversity of the learned Pareto sets. The visualizations further show that mPFM tends to provide more accurate approximations to the Pareto fronts on most benchmark problems, demonstrating the effectiveness of the proposed method.
Response (1/5)
Response to W1
To further evaluate our method in higher-dimensional decision spaces, we have conducted additional experiments on a 60-dimensional multiobjective Rover trajectory planning problem, extended from Wang et al. (2018) based on their source code https://github.com/zi-w/Ensemble-Bayesian-Optimization. The problem optimizes 30 free B-spline control points with two conflicting objectives: path length and obstacle risk. We further parameterize the obstacle configurations to construct related source tasks and consider four target settings, including two in-distribution (i.e., Rover-1 and Rover-3) and two out-of-distribution tasks (i.e., Rover-2 and Rover-4). We compare mPFM-MOBO with the five existing baselines over 20 independent runs, using the same budget of 100 evaluations (20 initial + 80 sequential evaluations).
(1) Hypervolume
(2) IGD+
(3) R2
As shown in the tables, mPFM-MOBO consistently achieves the best performance across all four target settings and all three metrics. Specifically, it achieves HV values of 1.144–1.206, substantially outperforming the strongest baseline, invTrEMO (0.758–0.958). Consistent improvements are also observed in IGD+ and R2. These results show that, for 60 dimensional problems, mPFM-MOBO also has a good performance.
Moreover, we clarify that our claim concerns improving Bayesian optimization's sample process, rather than reducing the computational difficulty of acquisition function optimization itself. By combining structured candidate generation with acquisition-based selection, the proposed method increases the likelihood of identifying high-quality solutions under limited evaluation budgets. We have updated the above results in the new uploaded manuscripts, please see Appendix I for details. We also have cited both recommended references [1, 2] in the revised appendix.