Pretrained Cost Model for Distributed Constraint Optimization Problems

Authors: Yanchen Deng, Shufeng Kong, Bo An9331-9340

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive empirical evaluations indicate that the GATPCM-boosted algorithms significantly outperform the stateof-the-art methods in various benchmarks.
Researcher Affiliation Academia Yanchen Deng, Shufeng Kong , Bo An School of Computer Science and Engineering, Nanyang Technological University, Singapore {ycdeng, shufeng.kong, boan}@ntu.edu.sg
Pseudocode Yes Algorithm 1: Offline pretraining procedure; Algorithm 2: Distributed embedding schema for agent i
Open Source Code Yes Our pretrained cost model is available at https://github.com/dyc941126/GAT-PCM.
Open Datasets No The paper describes how problem instances are generated for experiments (e.g., 'random DCOPs, scale-free networks, grid networks, and weighted graph coloring problems') and how data is sampled from distributions, but it does not provide access information (link, DOI, specific citation to a dataset) for a pre-existing publicly available dataset.
Dataset Splits No The paper describes the pretraining process and use of benchmarks, but it does not specify explicit training/validation/test dataset splits or percentages.
Hardware Specification Yes All experiments are conducted on an Intel i9-9820X workstation with Ge Force RTX 3090 GPUs.
Software Dependencies No Our model was implemented with the Py Torch Geometric framework (Fey and Lenssen 2019) and the model was trained with the Adam optimizer (Kingma and Ba 2014)... While PyTorch Geometric is mentioned, no specific version number for the framework or any other software dependencies is provided.
Experiment Setup Yes Our GAT-PCM model has four GAT layers (i.e., T = 4). Each layer in the first three layers has 8 output channels and 8 heads of attention, while the last layer has 16 output channels and 4 heads of attention. Each GAT layer uses ELU (Clevert, Unterthiner, and Hochreiter 2016) as the activation function. [...] For hyperparameters, we set the batch size and the number of training epochs to be 64 and 5000, respectively. [...] the model was trained with the Adam optimizer (Kingma and Ba 2014) using the learning rate of 0.0001 and a 5 × 10−5 weight decay ratio.