Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Neural Evolution Strategy for Black-box Pareto Set Learning
Authors: Chengyu Lu, Zhenhua Li, Xi Lin, Ji Cheng, Qingfu Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The proposed method, termed Neural-ES, is evaluated using a bespoke benchmark suite for black-box PSL. Experimental comparisons with other methods demonstrate the efficiency of Neural-ES, underscoring its ability to learn the Pareto sets of challenging black-box MOPs. |
| Researcher Affiliation | Academia | 1 City University of Hong Kong, 2 City University of Hong Kong Shenzhen Research Institute 3 Nanjing University of Aeronautics and Astronautics 4 MIIT Key Laboratory of Pattern Analysis and Machine Intelligence EMAIL, EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1: Neural-ES for black-box PSL |
| Open Source Code | Yes | Code of Neural-ES and the BBO-PSL suite is available in https://github.com/chandler09/ Neural-Evolution-Strategy. |
| Open Datasets | Yes | To this end, we present the novel BBO-PSL suite, which comprises eight instances (F1 to F8) and satisfies all the aforementioned requirements. ... Code of Neural-ES and the BBO-PSL suite is available in https://github.com/chandler09/ Neural-Evolution-Strategy. ... We consider a set of real-world multi-objective MO-UAV navigation problems, which are featured in the Meta BBO-v2 library [51] 3. ... 3https://github.com/Meta Evo/Meta Box |
| Dataset Splits | Yes | For the three PSL algorithms, we evaluate their testing performance. To do so, we input a large number of preferences into a trained set model, which yields the same number of distributions; from each distribution, we randomly sample a solution, the collection of which form the testing solution set; finally, we measure the IGD value of the set and refer it as the testing IGD value. Meanwhile, we track the convergence of the PSL algorithms during training, in terms of the IGD values acquired by the best solutions so far. ... The number of input preferences, i.e., the size of the testing solution set, is 900 and 4950 for the biand tri-objective instances, respectively. ... The training solution set at a certain iteration comprises the best solutions found so far for a set of evenly distributed subproblems. In light of the computational efficiency, its capacity is slightly smaller than the testing solution set, which is 100 and 300 for the biand tri-objective instances, respectively. |
| Hardware Specification | No | We record the information on the computer resources in the supplementary document. |
| Software Dependencies | No | The set model is trained using Adam, with a learning rate of η = 10 3. Compared to SGD (stochastic gradient ascent), the adaptive momentum of Adam accumulates historical gradient information, enabling more accurate gradient estimation and significantly reducing the number of samples required. |
| Experiment Setup | Yes | The set model is trained using Adam, with a learning rate of η = 10 3. ... configure Neural-ES by d = n/16 , N = 5n/16 , and λ = 2 + 1.5 ln n . Additionally, as MOEAs are widely recognized as one of the most effective methodologies for obtaining a set of non-dominated solutions, we consider two state-of-the-art MOEAs specifically designed for high-dimensional MOPs, namely LMO-CSO [35] and IM-MOEA/D [36]. ... The evaluation budget is 4000n and 2000n for biand tri-objective instances. |