BOtied: Multi-objective Bayesian optimization with tied multivariate ranks

Authors: Ji Won Park, Natasa Tagasovska, Michael Maser, Stephen Ra, Kyunghyun Cho

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on a variety of synthetic and real-world experiments demonstrate that BOtied outperforms state-of-the-art MOBO algorithms while being computationally efficient for many objectives.
Researcher Affiliation Collaboration 1Prescient Design, Genentech, South San Francisco, USA 2Department of Computer Science, New York University, New York City, USA 3Center for Data Science, New York University, New York City, USA.
Pseudocode Yes B. Algorithm Algorithm 1 MOBO with BOtied: a CDF-based acquisition function
Open Source Code Yes The code that reproduces all of our experiments and plots is available at https: //github.com/jiwoncpark/botied .
Open Datasets Yes We include synthetic test functions for direct evaluation of f. We focus on ones that support M 3: DTLZ2 (d=6, M=4 and d=7, M=6; Deb & Gupta, 2005) and Penicillin (d=7, M=3; Liang & Lai, 2021)... To emulate a multi-objective drug design setting, we postprocess the real-world dataset Caco2 (Wang et al., 2016) from the Therapeutics Data Commons database (Huang et al., 2021; 2022) to create Caco2+... We also include experiments over three datasets from the DDMOP benchmark (He et al., 2020).
Dataset Splits No The paper refers to an 'initial data size N0' and discusses the use of a probabilistic surrogate model (GP) but does not explicitly specify a validation dataset split or a methodology for creating one for reproduction purposes.
Hardware Specification Yes For all acquisition functions, we report the wall clock time per single acquisition function evaluation as computed on a Tesla V100 SXM2 GPU (16GB RAM) and an Intel Xeon CPU @ 2.30GHz (240GB RAM).
Software Dependencies No The paper mentions software like 'Bo Torch' and 'GPy Torch' but does not specify their version numbers or other key software dependencies with specific versions required for reproduction.
Experiment Setup Yes We executed batched BO simulations with sequential greedy optimization and varying batch sizes q {1, 2, 4}. The number of iterations T varied across the experiments... The initial data size N0, the size of the pool N, and the number of predictive posterior samples L. We fixed the size of the pool relative to the selected batch, at N/B = 100. We also fixed L = 20...