Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
A Deep Reinforcement Learning Framework for Column Generation
Authors: Cheng Chi, Amine Aboussalah, Elias Khalil, Juyoung Wang, Zoha Sherkat-Masoumi
NeurIPS 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform an extensive set of experiments using the publicly available BPPLIB benchmark for CSP and Solomon benchmark for VRPTW. RLCG converges faster and reduces the number of CG iterations by 22.4% for CSP and 40.9% for VRPTW on average compared to a commonly used greedy policy. |
| Researcher Affiliation | Academia | Cheng Chi University of Toronto Amine Mohamed Aboussalah New York University Elias B. Khalil University of Toronto Juyoung Wang University of Toronto Zoha Sherkat-Masoumi University of Toronto |
| Pseudocode | Yes | Algorithm 1 shows how a trained RLCG agent is applied to solve a problem instance. |
| Open Source Code | Yes | Our code is available at this link. |
| Open Datasets | Yes | We use BPPLIB Delorme et al. [2018], a widely used collection of benchmark instances for binary packing and cutting stock problems, which includes a number of instances proven to be difficult to solve [Delorme and Iori, 2020, Wei et al., 2020, Martinovic et al., 2020]. |
| Dataset Splits | Yes | Our training set has instances with n = 50, 100, 200. The remaining instances with n = 50, 100, 200, 750 are split into a validation set and a testing set, with hard instances n = 750 only appearing in the testing set, as it is very expensive to solve such large instances during training. |
| Hardware Specification | Yes | Our computing environment is described in Appendix Section C. |
| Software Dependencies | No | The paper does not explicitly state specific version numbers for software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | We tune the main hyperparameters, in the reward function (1), the exploration probability in DQN, γ the discount factor, and the learning rate lr in gradient descent. ... The best configuration is: = 300, = 0.05, γ = 0.9 and lr = 0.001. |