Offline Multi-Objective Optimization
Authors: Ke Xue, Rongxi Tan, Xiaobin Huang, Chao Qian
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results show improvements over the best value of the training set, demonstrating the effectiveness of offline MOO methods. [...] In this section, we empirically examine the performance of different methods on our benchmark. |
| Researcher Affiliation | Academia | 1National Key Laboratory for Novel Software Technology, Nanjing University, China 2School of Artificial Intelligence, Nanjing University, China. |
| Pseudocode | No | The paper describes its methods narratively and outlines network structures, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/ lamda-bbo/offline-moo. |
| Open Datasets | Yes | In this paper, we propose a first benchmark for offline MOO, where the tasks range from synthetic functions to real-world science and engineering problems... To facilitate future research, we release our benchmark tasks and datasets with a comprehensive evaluation of different approaches and open-source examples. [...] NAS-Bench-201-Test, corresponding error and number of parameters are sourced from Dong & Yang (2020). Additionally, the edge GPU latency data is obtained from Li et al. (2021). [...] We consider two locomotion tasks in the popular MORL benchmark Mu Jo Co (Todorov et al., 2012). [...] Historical stock prices data of each portfolio is provided by Blank & Deb (2020). [...] We also conduct experiments on seven real-world multi-objective engineering design problems adopted from RE suite (Tanabe & Ishibuchi, 2020). |
| Dataset Splits | Yes | Thus, similar to (Trabucco et al., 2022), we remove the top solutions sorted by NSGA-II ranking with a given percentile K, where K varies according to different tasks and is usually set 40%, except for Molecule with 1.2%, RFP and Regex with 20%, and MO-CVRP with 55%. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, memory, or cloud instance types used for running the experiments. |
| Software Dependencies | Yes | The implementations of NSGA-II, MOEA/D, and NSGA-III are from the open-source repository Py MOO (Blank & Deb, 2020). The implementation of MOBO is inherited from Bo Torch (Balandat et al., 2020). |
| Experiment Setup | Yes | The DNN model is trained w.r.t. offline dataset for 200 epochs with a batch size of 32. [...] We use MSE as loss function and optimize by Adam with learning rate η = 0.001 and learning-rate decay γ = 0.98. The model architecture and hyperparameters are consistently maintained across all tasks. |