Unsupervised Learning for Combinatorial Optimization with Principled Objective Relaxation

Authors: Haoyu Peter Wang, Nan Wu, Hang Yang, Cong Hao, Pan Li

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our framework by solving a synthetic graph optimization problem, and two real-world applications including resource allocation in circuit design and approximate computing. Our framework1 largely outperforms the baselines based on naïve relaxation, reinforcement learning, and Gumbel-softmax tricks.
Researcher Affiliation Academia Haoyu Wang Purdue University wang5272@purdue.edu Nan Wu UCSB nanwu@ucsb.edu Hang Yang Georgia Tech. hyang628@gatech.edu Cong Hao Georgia Tech. callie.hao@gatech.edu Pan Li Purdue University panli@purdue.edu
Pseudocode No The paper describes procedures (e.g., Definition 1 for rounding) in text, but no explicitly labeled pseudocode or algorithm blocks are present.
Open Source Code Yes Our code and the datasets are available at: https://github.com/Graph-COM/CO_Proxy Design
Open Datasets Yes Our code and the datasets are available at: https://github.com/Graph-COM/CO_Proxy Design
Dataset Splits Yes We follow the settings in [1] and use a 6:2:2 dataset split for training, validating and testing
Hardware Specification Yes To be fair, all methods run on the same server with a Quadro RTX 6000 GPU.
Software Dependencies No The paper mentions software like PyTorch [76], PyTorch Geometric [77], DGL [66], and the Adam optimizer [78], but does not provide specific version numbers for any of these dependencies.
Experiment Setup Yes The Adam optimizer [78] is adopted with a learning rate of 1e-4 and the batch size of 256. For RL training, we use a constant learning rate of 1e-5. We train both the proxies and the Aθ module for 200 epochs.