Adjusted Expected Improvement for Cumulative Regret Minimization in Noisy Bayesian Optimization

Authors: Shouri Hu, Haowei Wang, Zhongxiang Dai, Bryan Kian Hsiang Low, Szu Hui Ng

JMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We further perform experiments to illustrate the improvement of EIC over several popular BO algorithms. In the experiments, we examine the numerical performance of our EIC algorithm using synthetic test functions (Section 5.1) as well as a real-world dataset (Section 5.2). For each test function/dataset, we generate R = 100 independent experiment trials, and the results are summarized in Figures 2 and 3. The solid line represents the average cumulative regret over these independent trials, with the shaded area showing the corresponding 95% confidence region.
Researcher Affiliation Academia Shouri Hu EMAIL School of Mathematical Sciences University of Electronic Science and Technology of China, China; Haowei Wang EMAIL Department of Industrial Systems Engineering and Management National University of Singapore, Singapore; Zhongxiang Dai EMAIL School of Data Science The Chinese University of Hong Kong, Shenzhen, China; Bryan Kian Hsiang Low EMAIL Department of Computer Science National University of Singapore, Singapore; Szu Hui Ng EMAIL Department of Industrial Systems Engineering and Management National University of Singapore, Singapore
Pseudocode Yes Algorithm 1 Expected improvement-cost (EIC) algorithm Require: N, GP(µ, k), n0, ξn, ω2 n. Sample n0 initial design points as described in (11). Each point is sampled with one replication. for n = n0, . . . , N 1 do Update the GP posterior model GP(µn, kn) using the history of observations. if x D : αEI n (x) Ln(x) then Let Bn = {x D : αEI n (x) Ln(x)}. Select the point xn+1 = arg maxx:x Bn αEI n (x). else Select the point xn+1 = arg max1 i n µn(xi). end if Evaluate the selected point xn+1 with one replication and observe yn+1. Add the newly collected (xn+1, yn+1) to the history of observations. end for
Open Source Code No The paper does not provide a concrete statement about open-sourcing the code for the methodology described in this paper, nor does it provide a link to a code repository.
Open Datasets Yes We adopt the breast cancer Wisconsin dataset (https://archive.ics.uci.edu/ml/datasets/breast+cancer+wisconsin+(diagnostic)) which contains 569 patients.
Dataset Splits Yes The dataset is further split into training and testing datasets with a ratio of 7:3.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions general techniques and models (e.g., GP prior, SE covariance kernel, neural network model) and cites a method for parameter estimation, but it does not specify any software names with version numbers that would be needed to replicate the experiment.
Experiment Setup Yes The total budget N for each test function/dataset is set to N = 200 + n0, where the number of initial design points is set at n0 = 16, 36, 64 for the 2, 4, 6 dimensional objective functions respectively. The standard deviation of the homogenous noise σ is set as 0.1. We consider the tuning of four hyper-parameters: the number of hidden units, batch size, learning rate, and learning rate decay coefficients. We follow their suggestion and choose κ = 10 4 in our experiments.