Simulation-guided Beam Search for Neural Combinatorial Optimization

Authors: Jinho Choo, Yeong-Dae Kwon, Jihoon Kim, Jeongwoo Jae, André Hottung, Kevin Tierney, Youngjune Gwon

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our methods on well-known CO benchmarks and show that SGBS significantly improves the quality of the solutions found under reasonable runtime assumptions. We show the effectiveness both of SGBS and the SGBS+EAS hybrid on a standard benchmark of the traveling salesperson problem (TSP), capacitated vehicle routing problem (CVRP), and the flexible flow shop problem (FFSP).
Researcher Affiliation Collaboration Jinho Choo 1 Yeong-Dae Kwon 1 Jihoon Kim1 Jeongwoo Jae1 Andr e Hottung2 Kevin Tierney2 Youngjune Gwon1 1Samsung SDS, Korea 2Bielefeld University, Germany
Pseudocode Yes Algorithm 1 Simulation-guided Beam Search (SGBS)
Open Source Code Yes Our code for the experiments described in the paper is publicly available at https://github.com/ yd-kwon/SGBS.
Open Datasets Yes For our experiments, we use n = 100 with 10,000 instances from Kool et al. [4].
Dataset Splits No The paper states that models were 'pre-trained by the POMO [14] RL technique' and uses 'n = 100 with 10,000 instances' and 'n = 150, 200 test sets of 1,000 instances' for evaluation, but does not explicitly provide the specific train/validation/test splits used for the training process itself.
Hardware Specification Yes Our experiments are carried out on A100 GPUs (Nvidia) with 80 GB memory.
Software Dependencies Yes Results of our SGBS experiments and the ablation studies are summarized in Table 2 along with the results [13] by CPLEX [51] with mixed-integer programming models and meta-heuristic solvers.
Experiment Setup Yes SGBS hyperparameters are set to (β, γ) = (10, 10) and (4, 4) for the TSP and CVRP, respectively.