Efficient Neural Neighborhood Search for Pickup and Delivery Problems

Authors: Yining Ma, Jingwen Li, Zhiguang Cao, Wen Song, Hongliang Guo, Yuejiao Gong, Yeow Meng Chee

IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our N2S is generic, and extensive experiments on two canonical PDP variants show that it can produce state-of-the-art results among existing neural methods. Moreover, it even outstrips the well-known LKH3 solver on the more constrained PDP variant.
Researcher Affiliation Collaboration 1National University of Singapore 2Singapore Institute of Manufacturing Technology, A*STAR 3Institute of Marine Science and Technology, Shandong University 4Institute for Infocomm Research, A*STAR 5South China University of Technology
Pseudocode Yes Algorithm 1 n-step PPO with CL strategy Algorithm 2 N2S-A Inference
Open Source Code Yes Our implementation for N2S is available online1. 1Code is available at https://github.com/yining043/PDP-N2S.
Open Datasets Yes We evaluate N2S on PDTSP and PDTSP-LIFO... where the node coordinates of instances are randomly and uniformly generated in the unit square [0, 1] [0, 1]... We further evaluate our N2S on benchmark instances, including all the ones from [Renaud et al., 2002] for PDTSP and the ones with size |V| ≥ 201 from [Carrabs et al., 2007] for PDTSP-LIFO
Dataset Splits No The paper specifies training and test datasets ('All baselines are evaluated on a test dataset with 2,000 instances'), but does not explicitly mention a separate validation dataset split or its size/proportion in the main text.
Hardware Specification Yes Our experiments were conducted on a server equipped with 8 RTX 2080 Ti GPU cards and Intel E5-2680 CPU @ 2.4GHz.
Software Dependencies No The paper mentions 'Adam optimizer' but does not specify version numbers for any programming languages, libraries, or other software components used in the experiments.
Experiment Setup Yes Our N2S is trained with E = 200 epochs and B = 20 batches per epoch using batch size 600. We set n=5, Ttrain =250 for the n-step PPO with κ=3 mini-batch updates and a clip threshold ϵ=0.1. Adam optimizer is used with learning rate ηθ =8·10−5 for πθ, and ηϕ =2·10−5 for vϕ (decayed β = 0.985 per epoch). The discount factor γ is set to 0.999 for both PDPs. We clip the gradient norm to be within 0.05, 0.15, 0.35, and set the curriculum learning ρCL to 2, 1.5, 1 for the three problem sizes, respectively.