TY - GEN
T1 - ClassBO
T2 - 18th International Conference on Learning and Intelligent Optimization, LION 2024
AU - Malu, Mohit
AU - Pedrielli, Giulia
AU - Dasarathy, Gautam
AU - Spanias, Andreas
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
PY - 2025
Y1 - 2025
N2 - Bayesian Optimization (BO) frameworks typically assume the function to be optimized is stationary (homogeneous) over the domain. However, in many real-world applications, we often deal with functions that present a rate of variation across the input space. In this paper, we optimize functions where a finite set of homogeneous functions defined over partitions of the input space can represent the heterogeneity. The disconnected partitions that can be characterized by the same function are said to be in the same class, and evaluating the function at input returns the minimum distance to a boundary of the contiguous class (partition). The ClassGP modeling framework, previously developed to model for such heterogenous functions along with a novel ClassUCB acquisition function and partition sampling strategy, is used to introduce a novel tree-based optimization framework dubbed as ClassBO (Class Bayesian Optimization). We demonstrate the superior performance of ClassBO against other methods via empirical evaluations.
AB - Bayesian Optimization (BO) frameworks typically assume the function to be optimized is stationary (homogeneous) over the domain. However, in many real-world applications, we often deal with functions that present a rate of variation across the input space. In this paper, we optimize functions where a finite set of homogeneous functions defined over partitions of the input space can represent the heterogeneity. The disconnected partitions that can be characterized by the same function are said to be in the same class, and evaluating the function at input returns the minimum distance to a boundary of the contiguous class (partition). The ClassGP modeling framework, previously developed to model for such heterogenous functions along with a novel ClassUCB acquisition function and partition sampling strategy, is used to introduce a novel tree-based optimization framework dubbed as ClassBO (Class Bayesian Optimization). We demonstrate the superior performance of ClassBO against other methods via empirical evaluations.
KW - Bayesian Optimization
KW - Black-box Optimization
KW - Gaussian process
KW - Heterogeneous function
KW - Non-stationary function
UR - http://www.scopus.com/inward/record.url?scp=85216123083&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85216123083&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-75623-8_19
DO - 10.1007/978-3-031-75623-8_19
M3 - Conference contribution
AN - SCOPUS:85216123083
SN - 9783031756221
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 249
EP - 253
BT - Learning and Intelligent Optimization - 18th International Conference, LION 18, Revised Selected Papers
A2 - Festa, Paola
A2 - Ferone, Daniele
A2 - Pastore, Tommaso
A2 - Pisacane, Ornella
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 9 June 2024 through 13 June 2024
ER -