Uncertainty-Aware Search Framework for Multi-Objective Bayesian Optimization

Authors: Syrine Belakaria, Aryan Deshwal, Nitthilan Kannappan Jayakodi, Janardhan Rao Doppa10044-10052

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on several synthetic and six diverse real-world benchmark problems show that USe MO consistently outperforms the state-of-the-art algorithms.
Researcher Affiliation Academia Syrine Belakaria, Aryan Deshwal, Nitthilan Kannappan Jayakodi, Janardhan Rao Doppa School of EECS, Washington State University {syrine.belakaria, aryan.deshwal, n.kannappanjayakodi, jana.doppa}@wsu.edu
Pseudocode Yes Algorithm 1 USe MO Framework
Open Source Code No The paper mentions using 'the code for these methods from the BO library Spearmint1. 1https://github.com/HIPS/Spearmint/tree/PESM' for existing methods, but it does not provide an explicit statement or link for the open-source code of their proposed USe MO framework.
Open Datasets Yes We optimize a dense neural network over the MNIST dataset (Le Cun et al. 1998)...SW-LLVM is a data set with 1024 compiler settings (Siegmund et al. 2012)...The data set SNW was first introduced by (Zuluaga, Milder, and P uschel 2012)...The design space of No C dataset (Almer, Topham, and Franke 2011)...The materials dataset SMA consists of 77 different design configu-rations of shape memory alloys (Gopakumar et al. 2018)...PEM is a materials dataset consisting of 704 configurations of Piezoelectric materials (Gopakumar et al. 2018).
Dataset Splits Yes We employ 10K instances for validation and 50K instances for training.
Hardware Specification Yes we run all experiments on a machine with the following configuration: Intel i7-7700K CPU @ 4.20GHz with 8 cores and 32 GB memory.
Software Dependencies No The paper mentions using 'GP based statistical model', 'NSGA-II algorithm', and the 'BO library Spearmint', but it does not provide specific version numbers for any of these software components.
Experiment Setup Yes The hyper-parameters are estimated after every 10 function evaluations...We initialize the GP models for all functions by sampling initial points at random from a Sobol grid using the in-built procedure in the Spearmint library...For NSGA-II, the most important parameter is the number of function calls...Therefore, we fixed it to 1500 for all our experiments...We train the network for 100 epochs for evaluating each candidate hyper-parameter values on validation set.