Neural Multi-Objective Combinatorial Optimization via Graph-Image Multimodal Fusion

Authors: Jinbiao Chen, Jiahai Wang, Zhiguang Cao, Yaoxin Wu

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on classic MOCO problems show that our GIMF significantly outperforms state-of-the-art neural MOCO methods and exhibits superior generalization capability.
Researcher Affiliation Academia 1School of Computer Science and Engineering, Sun Yat-sen University, P.R. China 2Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, Sun Yat-sen University, P.R. China 3Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou, P.R. China 4School of Computing and Information Systems, Singapore Management University, Singapore 5Department of Industrial Engineering & Innovation Sciences, Eindhoven University of Technology
Pseudocode No The paper describes the methodologies and processes using prose and mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code is publicly available1. 1https://github.com/bill-cjb/GIMF
Open Datasets Yes The instances are generated by uniformly sampling coordinates within the range [0, 1]2M.
Dataset Splits No The paper describes using randomly selected instances for training and evaluating on 200 random instances or benchmark instances, but does not provide specific training/test/validation dataset splits (percentages or counts).
Hardware Specification Yes All methods adopt WS scalarization for fair comparisons, and are executed on a machine equipped with an RTX 3090 GPU and an Intel Xeon 4216 CPU.
Software Dependencies No The paper mentions using the Adam optimizer, but does not specify any software libraries or their version numbers (e.g., PyTorch, TensorFlow, Python version, CUDA version) used for the implementation.
Experiment Setup Yes Most hyperparameters for GIMF-P and GIMF-C are configured in line with the original PMOCO and CNH, respectively. For our model, L = 6, L = 3, and nb = n b = 10. The patch dimensions are fixed at w = h = 16. The model undergoes training for 200 epochs, with each epoch processing 100,000 randomly selected instances and a batch size of B = 64. The Adam optimizer (Kingma & Ba, 2015) is used with a learning rate of 10 4 (except 10 5 for Bi-KP) and weight decay of 10 6. The N weight vectors for the decomposition are generated according to Das & Dennis (1998), with N = 101 for M = 2 and N = 105 for M = 3.