Data-Driven Based State Transition Algorithm for Dynamic Optimization

Most evolutionary algorithms solve the optimization problems by iteratively calculating the values of the objective functions. However, such explicit objective functions may not exist in solving many process industry optimization problems, especially when these processes have time varying dynamics,...

Full description

Saved in:
Bibliographic Details
Published in:2019 IEEE Symposium Series on Computational Intelligence (SSCI) pp. 1091 - 1096
Main Authors: Zhang, Yunxiang, Zhou, Xiaojun, Yang, Chunhua
Format: Conference Proceeding
Language:English
Published: IEEE 01.12.2019
Subjects:
Approximation algorithms
Control vector parametrization
Data-driven
Dynamic optimization
Evolutionary computation
Heuristic algorithms
Linear programming
Optimization
Process control
State transition algorithm
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Most evolutionary algorithms solve the optimization problems by iteratively calculating the values of the objective functions. However, such explicit objective functions may not exist in solving many process industry optimization problems, especially when these processes have time varying dynamics, instead costly physical experiments should be conducted to evaluate the fitness values, which hinders the application of evolutionary algorithms. In this paper, a novel dynamic optimization technique based on data-driven state transition algorithm (STA) is investigated to solve the aforementioned issue. Firstly, the control vector parameterization (CVP) method is used to discretize the control variables, and control variable vectors are used to approximate the control trajectories. By means of costly experiments, data pairs are obtained by combining variable vectors and their corresponding fitness. Then, a novel method is proposed to obtain the optimal variable vector based on data-driven STA. By applying it in two well-known dynamic optimization instances, simulation results demonstrate that the proposed approach is able to obtain better optimization results than other methods with a limited budget of exact function evaluations.
DOI:10.1109/SSCI44817.2019.9003081