Bayesian Optimization of Combinatorial Structures

Authors: Ricardo Baptista, Matthias Poloczek

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental evaluations demonstrate that this algorithm consistently outperforms other methods from combinatorial and Bayesian optimization.
Researcher Affiliation Academia 1Center for Computational Engineering, Massachusetts Institute of Technology, Cambridge, MA 2Dept. of Systems and Industrial Engineering, The University of Arizona, Tucson, AZ.
Pseudocode Yes BOCS is summarized as Algorithm 1. Algorithm 1 Bayesian Optimization of Combinatorial Structures
Open Source Code Yes The code for this paper is available at https://github. com/baptistar/BOCS.
Open Datasets Yes We conduct experiments on the following benchmarks: (1) binary quadratic programming with d = 10 variables (Sect. 4.1), (2) sparsification of Ising models with d = 24 edges (Sect. 4.2), (3) contamination control of a food supply chain with d = 25 stages (Sect. 4.3), (4) complexity reduction of an aero-structural multi-component system with d = 21 coupling variables (Sect. 4.4). We consider zero-field Ising models that admit a distribution p(z) = 1 Zp exp(z T Jpz) for z { 1, 1}n, where Jp Rn n is a symmetric interaction matrix and Zp is the partition function. The contamination control problem (Hu et al., 2010) considers a food supply with d stages that may be contaminated with pathogenic microorganisms. We study the aero-structural problem of Jasa et al. (2018).
Dataset Splits No The paper does not provide specific training, validation, and test dataset splits. It describes initial sample sizes for its optimization problem.
Hardware Specification No The paper does not specify the hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions that code is available and refers to the use of an SDP solver, but it does not list specific software dependencies with version numbers.
Experiment Setup Yes We set d = 10, sampled 50 independent realizations for Q, and ran every algorithm 10-times on each instance with different initial datasets. Bayesian optimization algorithms received identical initial datasets of size N0 = 20. The experimental setup consists of 4 4 zero-field Ising models with grid graphs, i.e., n = 16 nodes and d = 24 edges. The initial dataset contains N0 = 20 points. We have d = 25 stages. We set T = 102, hence the objective requires T simulations of the random process and is expensive to evaluate. The ℓ1 regularization term encourages the prevention efforts to occur at a small number of stages. Fig. 6 shows the average performances for λ = 10 2.