IPAC2021 - List of Authors (Hidas, D.A.)

Paper	Title	Page
THPAB197	Enhancing Efficiency of Multi-Objective Neural-Network-Assisted Nonlinear Dynamics Lattice Optimization via 1-D Aperture Objectives & Objective Focusing	4156
	Y. Hidaka, D.A. Hidas, F. Plassard, T.V. Shaftan, G.M. Wang BNL, Upton, New York, USA
	Funding: This work is supported by U.S. DOE under Contract No. DE-SC0012704. Mutli-objective optimizers such as multi-objective genetic algorithm (MOGA) have been quite popular in discovering desirable lattice solutions for accelerators. However, even these successful algorithms can become ineffective as the dimension and range of the search space increase due to exponential growth in the amount of exploration required to find global optima. This difficulty is even more exacerbated by the resource-intensive and time-consuming tendency for the evaluations of nonlinear beam dynamics. Lately the use of surrogate models based on neural network has been drawing attention to alleviate this problem. Following this trend, to further enhance the efficiency of nonlinear lattice optimization for storage rings, we propose to replace typically used objectives with those that are less time-consuming and to focus on a single objective constructed from multiple objectives, which can maximize utilization of the trained models through local optimization and objective gradient extraction. We demonstrate these enhancements using a NSLS-II upgrade lattice candidate as an example.
DOI •	reference for this paper ※ https://doi.org/10.18429/JACoW-IPAC2021-THPAB197
About •	paper received ※ 20 May 2021 paper accepted ※ 23 June 2021 issue date ※ 10 August 2021
Export •	reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)

Paper

Title

Page

Enhancing Efficiency of Multi-Objective Neural-Network-Assisted Nonlinear Dynamics Lattice Optimization via 1-D Aperture Objectives & Objective Focusing

4156

Y. Hidaka, D.A. Hidas, F. Plassard, T.V. Shaftan, G.M. Wang
BNL, Upton, New York, USA

Funding: This work is supported by U.S. DOE under Contract No. DE-SC0012704.
Mutli-objective optimizers such as multi-objective genetic algorithm (MOGA) have been quite popular in discovering desirable lattice solutions for accelerators. However, even these successful algorithms can become ineffective as the dimension and range of the search space increase due to exponential growth in the amount of exploration required to find global optima. This difficulty is even more exacerbated by the resource-intensive and time-consuming tendency for the evaluations of nonlinear beam dynamics. Lately the use of surrogate models based on neural network has been drawing attention to alleviate this problem. Following this trend, to further enhance the efficiency of nonlinear lattice optimization for storage rings, we propose to replace typically used objectives with those that are less time-consuming and to focus on a single objective constructed from multiple objectives, which can maximize utilization of the trained models through local optimization and objective gradient extraction. We demonstrate these enhancements using a NSLS-II upgrade lattice candidate as an example.

DOI •

reference for this paper ※ https://doi.org/10.18429/JACoW-IPAC2021-THPAB197

About •

paper received ※ 20 May 2021 paper accepted ※ 23 June 2021 issue date ※ 10 August 2021

Export •

reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)