Neural Multi-Objective Combinatorial Optimization with Diversity Enhancement

Authors: Jinbiao Chen, Zizhen Zhang, Zhiguang Cao, Yaoxin Wu, Yining Ma, Te Ye, Jiahai Wang

NeurIPS 2023 | Conference 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 NHDE is able to generate a Pareto front with higher diversity, thereby achieving superior overall performance.
Researcher Affiliation Academia Jinbiao Chen1, Zizhen Zhang1, Zhiguang Cao2, Yaoxin Wu3, Yining Ma4, Te Ye1, and Jiahai Wang1,5,6, 1School of Computer Science and Engineering, Sun Yat-sen University, P.R. China 2School of Computing and Information Systems, Singapore Management University, Singapore 3Department of Industrial Engineering & Innovation Sciences, Eindhoven University of Technology, Netherlands 4Department of Industrial Systems Engineering & Management, National University of Singapore, Singapore 5Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, Sun Yat-sen University, P.R. China 6Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou, P.R. China
Pseudocode Yes Algorithm 1 Training algorithm of NHDE-P
Open Source Code Yes Our code is publicly available3. 3https://github.com/bill-cjb/NHDE
Open Datasets Yes We evaluate the proposed NHDE on three typical MOCO problems that are commonly studied in the neural MOCO literature [13 15], namely the multi-objective traveling salesman problem (MOTSP) [51], multi-objective capacitated vehicle routing problem (MOCVRP) [3], and multi-objective knapsack problem (MOKP) [52]. ... Three commonly used benchmark instances developed from TSPLIB [56], i.e., Kro AB100, Kro AB150, and Kro AB200, are also tested.
Dataset Splits No The paper mentions training on randomly generated instances and evaluating on test instances, but does not specify a distinct validation split or its size/percentage.
Hardware Specification Yes All the methods are tested with an RTX 3090 GPU and an Intel Xeon 4216 CPU.
Software Dependencies No The paper mentions using the Adam optimizer and that PPLS/D-C is 'implemented in Python', but it does not provide specific version numbers for any software components (e.g., Python, PyTorch, CUDA, specific library versions).
Experiment Setup Yes We train NHDE-P with 200 epochs, each containing 5,000 randomly generated instances. We use batch size B =64 and the Adam [53] optimizer with learning rate 10 4 (10 5 for MOKP) and weight decay 10 6. During training, N =20 weights are sampled for each instance. During inference, we generate N = 40 and N = 210 uniformly distributed weights for M = 2 and M =3, respectively, which are then shuffled so as to counteract biases. The diversity factors linearly shift through the N subproblems from (1,0) to (0,1), which implies a gradual focus from achieving convergence (scalar objective) with a few solutions to ensuring comprehensive performance with a multitude of solutions. We set K =20 and J =200.