Bézier Flow: a Surface-wise Gradient Descent Method for Multi-objective Optimization

Authors: Akiyoshi Sannai, Yasunari Hikima, Ken Kobayashi, Akinori Tanaka, Naoki Hamada

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conducted numerical experiments with synthetic and real multi-objective optimization problem instances and demonstrated that our method achieved lower generalization errors than the existing multi-objective optimization algorithms.
Researcher Affiliation Collaboration Akiyoshi Sannai EMAIL Kyoto University National Institute of Informatics Yasunari Hikima EMAIL Fujitsu Limited Kyushu University Ken Kobayashi EMAIL Institute of Science Tokyo Akinori Tanaka EMAIL RIKEN AIP RIKEN i THEMS Keio University Naoki Hamada EMAIL KLab Inc.
Pseudocode Yes Algorithm 1 Multi-objective Optimization Method M(A) from Single Optimization Method A Algorithm 2 Surface-wise Gradient Descent Method
Open Source Code Yes The source code is accessible at https://github.com/hikimay/bezier-flow.
Open Datasets Yes For real-world instances, we applied Algorithm 2 to a group LASSO problem with the Birthwt dataset from the MASS package1 in R language. In addition, to confirm the applicability of the proposed method to non-simplicial instances, we applied Algorithm 2 to DTLZ instances (Deb et al., 2002)
Dataset Splits Yes As a validation set Y , we generated approximate Pareto solutions by NSGA-III (Deb & Jain, 2013; Jain & Deb, 2013) with the population size of 1000. To construct X, we randomly sampled {ˆtn}1000 n=1 i.i.d. from the uniform distribution on M−1 and obtained sample points on the estimated Bézier simplex.
Hardware Specification Yes the experiments were performed on a Windows 10 PC with an Intel(R) Xeon(R) W-1270 CPU 3.40 GHz and 64 GB RAM.
Software Dependencies Yes We implemented these algorithms in Python 3.12.3
Experiment Setup Yes In Algorithm 2, we set the degree of Bézier simplex as D = 3, the initial control points as P (1) = O, which is the zero matrix of appropriate size. In addition, we set the maximum number of iterations as K = 100 and the step size as α(k) = 1/k for k = 1, 2, . . . , K. The number of points to be sampled from a simplex in each iteration was tested for N {30, 50, 100}.