Sampling with Riemannian Hamiltonian Monte Carlo in a Constrained Space
Authors: Yunbum Kook, Yin-Tat Lee, Ruoqi Shen, Santosh Vempala
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate for the first time that ill-conditioned, non-smooth, constrained distributions in very high dimension, upwards of 100,000, can be sampled efficiently in practice. ... On benchmark data sets from systems biology and linear programming, our algorithm outperforms existing packages by orders of magnitude. ... In this section, we demonstrate the efficiency of our sampler using experiments on real-world datasets and compare our sampler with existing samplers. |
| Researcher Affiliation | Collaboration | Yunbum Kook Georgia Tech yb.kook@gatech.edu Yin Tat Lee Microsoft Research, University of Washington yintat@uw.edu Ruoqi Shen University of Washington shenr3@cs.washington.edu Santosh S. Vempala Georgia Tech vempala@gatech.edu |
| Pseudocode | Yes | Algorithm 1: Riemannian Hamiltonian Monte Carlo (RHMC) |
| Open Source Code | Yes | Our complete package is available on Git Hub. ... In the experiments, we used our MATLAB and C++ implementation of CRHMC3, which is available here and has been integrated into the COBRA toolbox. ... We provide an anonymized link to our algorithm package in Section 3.1. |
| Open Datasets | Yes | We used twelve constraint-based metabolic models from molecular systems biology in the COBRA Toolbox v3.0 [21] and ten real-world LP examples randomly chosen from NETLIB LP test sets. |
| Dataset Splits | No | Our algorithm does not require training. |
| Hardware Specification | Yes | Settings. We performed experiments on the Standard DS12 v2 model from MS Azure cloud, which has a 2.1GHz Intel Xeon Platinum 8171M CPU and 28GB memory. |
| Software Dependencies | No | In the experiments, we used our MATLAB and C++ implementation of CRHMC3... We used twelve constraint-based metabolic models from molecular systems biology in the COBRA Toolbox v3.0 [21]... We used as a baseline the Coordinate Hit-and-Run (CHAR) implemented in two different languages... the same algorithm (CDHR) with an R interface and a C++ library, Vol Esti [6]. While COBRA Toolbox and Vol Esti have versions, general software like MATLAB and C++ compilers are not specified with versions, nor are other potential system dependencies. |
| Experiment Setup | Yes | Each algorithm attempted to draw 1000 uniform samples, with limits on running time set to 1 day (3 days for the largest instance ken_18) and memory usage to 6GB. |