TY - GEN
T1 - A New Acquisition Function for Multi-objective Bayesian Optimization
T2 - 2023 Genetic and Evolutionary Computation Conference Companion, GECCO 2023 Companion
AU - Yang, Kaifeng
AU - Chen, Kai
AU - Affenzeller, Michael
AU - Werth, Bernhard
N1 - Funding Information:
This work is supported by the Austrian Science Fund (FWF – Der Wissenschaftsfonds) under the project (I 5315, ‘ML Methods for Feature Identification Global Optimization). This work is partly supported by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence, The Chinese University of Hong Kong, Shenzhen, under Grant No. 2022B1212010001, by the Natural Science Foundation of China (NSFC) with grant No. 62106212, and by the High Performance Computing Center of Central South University (CSU), China.
Publisher Copyright:
© 2023 Copyright held by the owner/author(s).
PY - 2023/7/15
Y1 - 2023/7/15
N2 - Multi-objective Bayesian optimization is a sequential optimization strategy in which an optimizer searches for optimal solutions by maximizing an acquisition function. Most existing acquisition functions assume that objectives are independent, but none of them incorporates the correlations among objectives through an explicit formula for exact computation. This paper proposes a novel acquisition function, namely, correlated probability of improvement (cPoI), for bi-objective optimization problems. The cPoI method builds on the probability of improvement and addresses the correlations between objectives by utilizing 3 distinct approaches to compute the posterior covariance matrix from a multi-task Gaussian process. This paper presents both an explicit formula for exact computation of cPoI and a Monte Carlo method for approximating it. We evaluate the performance of the proposed cPoI against 4 state-of-the-art multi-objective optimization algorithms on 8 artificial benchmarks and 1 real-world problem. Our experimental results demonstrate the effectiveness of cPoI in achieving superior optimization performance.
AB - Multi-objective Bayesian optimization is a sequential optimization strategy in which an optimizer searches for optimal solutions by maximizing an acquisition function. Most existing acquisition functions assume that objectives are independent, but none of them incorporates the correlations among objectives through an explicit formula for exact computation. This paper proposes a novel acquisition function, namely, correlated probability of improvement (cPoI), for bi-objective optimization problems. The cPoI method builds on the probability of improvement and addresses the correlations between objectives by utilizing 3 distinct approaches to compute the posterior covariance matrix from a multi-task Gaussian process. This paper presents both an explicit formula for exact computation of cPoI and a Monte Carlo method for approximating it. We evaluate the performance of the proposed cPoI against 4 state-of-the-art multi-objective optimization algorithms on 8 artificial benchmarks and 1 real-world problem. Our experimental results demonstrate the effectiveness of cPoI in achieving superior optimization performance.
KW - Multi-objective Bayesian Optimization
KW - Multi-task Gaussian Process
KW - Posterior Covariance
KW - Probability of Improvement
UR - http://www.scopus.com/inward/record.url?scp=85169018189&partnerID=8YFLogxK
U2 - 10.1145/3583133.3596374
DO - 10.1145/3583133.3596374
M3 - Conference contribution
AN - SCOPUS:85169018189
T3 - GECCO 2023 Companion - Proceedings of the 2023 Genetic and Evolutionary Computation Conference Companion
SP - 2308
EP - 2317
BT - GECCO 2023 Companion - Proceedings of the 2023 Genetic and Evolutionary Computation Conference Companion
PB - Association for Computing Machinery, Inc
Y2 - 15 July 2023 through 19 July 2023
ER -