Triangulation candidates for Bayesian optimization
Authors: Robert B. Gramacy, Annie Sauer, Nathan Wycoff
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Here we propose using candidates based on a Delaunay triangulation of the existing input design. We detail the construction of these tricands and demonstrate empirically how they outperform both numerically optimized acquisitions and random candidate-based alternatives, and are well-suited for hybrid schemes, on benchmark synthetic and real simulation experiments. |
| Researcher Affiliation | Academia | Robert B. Gramacy Department of Statistics Virginia Tech Blacksburg, VA 24060, USA Annie Sauer Department of Statistics Virginia Tech Blacksburg, VA 24060, USA anniees@vt.edu Nathan Wycoff The Mc Court School s Massive Data Institute Georgetown University Washington DC, 20057, USA nathan.wycoff@georgetown.edu |
| Pseudocode | No | The paper describes computational procedures in detail, but it does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our implementation, provided for Python and R in our git repository,1 is relatively tidy. For example, tricands.R therein contains just 71 lines of code, eleven of which are to support optional visualizations in 2d such as those in Figure 2. The heavy lifting is done by Qhull. |
| Open Datasets | Yes | Our synthetic fs, including those in Section 4, are described in more detail on the pages of the Virtual Library for Simulation Experiments [VLSE; Surjanovic and Bingham, 2013]. |
| Dataset Splits | No | The paper describes the initial design size (n0 = 12 or n0 = 60) and the total number of acquisitions (nend = 50 or nend = 300) for sequential optimization experiments. However, it does not specify fixed training, validation, and test dataset splits as typically found in traditional machine learning experiments. |
| Hardware Specification | Yes | using a single core of an Intel i7-6900K CPU at 3.20GHz. |
| Software Dependencies | No | The paper mentions several software packages and libraries (e.g., Python, R, scipy.spatial, Qhull, tgp on CRAN, deepgp on CRAN). However, it does not provide specific version numbers for all key software components (e.g., Python, scipy, Qhull, tgp, deepgp). It mentions 'R package version 3.6.2' for R.matlab, but this is an ancillary tool, not a core dependency for the main methodology. |
| Experiment Setup | Yes | We consider three methods for solving EI acquisitions: a continuous search of the criterion via L-BFGS-B with 5-multi-starts, LHS candidates, and tricands. For TS acquisitions, we similarly employ both LHS candidates and tricands. As non-BO (not surrogate-based) comparators, serving primarily as benchmarks, we entertain raw L-BFGS-B and Nelder Mead [Nelder and Mead, 1965]. |