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 and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Neural Combinatorial Optimization for Time-Dependent Traveling Salesman Problem
Authors: Ruixiao Yang, Chuchu Fan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Instead, we present extensive experiments on real-world datasets, evaluating our approach on both entire datasets and specifically filtered instances where temporal dependencies alter the optimal solution. Results show that our method achieves state-of-the-art average optimality gap on full instances and significant travel-time reduction on instances where time-aware routing saves time. These results demonstrate state-of-the-art ability to identify and exploit temporal dependencies, setting new standards for evaluating time-dependent routing problems. |
| Researcher Affiliation | Academia | Ruixiao Yang Massachusetts Institute of Technology EMAIL Chuchu Fan Massachusetts Institute of Technology EMAIL |
| Pseudocode | Yes | Algorithm 1 Greedy Algorithm Input: node set V , start node π1, start time t1, time dependent cost function c Output: A permutation π of the node set V as a TDTSP tour. 1: for i = 2, 3, , n do 2: πi = arg minv V \π[1:i 1] cπi 1,v(tπi 1) 3: ti = ti 1 + cπi 1,πi(tπi 1) 4: end for |
| Open Source Code | Yes | We also provide code in the supplementary material. |
| Open Datasets | Yes | We generate data from real-world datasets across four cities: Beijing [38], Lyon [26], Nairobi, and London [5]. These datasets contain time-dependent travel times between nodes, with varying sampling intervals and time horizons. |
| Dataset Splits | Yes | From each city s data, we generate problem instances with n {10, 20} nodes via downsampling (50-node instances were generated only for Beijing due to time-horizon constraints in the other cities). For each problem size and city combination, we generated 10,000 test instances. |
| Hardware Specification | Yes | All the experiments were conducted on a workstation with 128 Ryzen Threadripper PRO 7985WX 64-Cores CPU and 4 NVIDIA A800 GPUs. |
| Software Dependencies | No | Our proposed approach was programmed with Py Torch. |
| Experiment Setup | Yes | We train our model with a batch size of 1024 for instances of size 10 and 20, and 256 for instances of size 50. We use Adam optimizer with a learning rate 1e-4. We train 200, 300, and 500 epochs, respectively, with each epoch 1,280,000 data. |