TY - GEN
T1 - Feature selection in evolutionary algorithm-based parameter estimation of duffing oscillators
AU - Banerjee, Amit
AU - Mahfouz, Issam Abu
AU - Abu-Ayyad, Ma'moun
PY - 2013
Y1 - 2013
N2 - The use of evolutionary optimization techniques such as genetic algorithms, differential evolution, swarm optimization and genetic programming to solve the inverse problem of parameter estimation for nonlinear chaotic systems has been gaining popularity in recent years. The efficacy of such evolutionary schemes depends on the definition of a suitable fitness function which is used to compare potential solutions in the population. In almost all research involving evolutionary schemes for parameter identification, displacement values of the first few hundred Poincaré points, after ignoring transient effects, have been used as the feature set. The measured response of the system is compared to the response of the potential solutions in the population over these Poincaré points, although there is no empirical research to show that such a feature set works better than other possible feature sets. In this paper, a smaller feature set based on first and second-order statistical parameters of the response are considered and the estimation results are compared to the estimate produced by using the standard Poincaré points-based feature set, called the finite sample feature set in this paper. Also compared are results using three evolutionary algorithms - firefly algorithm, particle swarm optimization and differential evolution. It has been shown that the proposed feature set converges to a near-optimal solution faster and in fewer generations and produces estimates that are comparable to those obtained with the finite sample feature set.
AB - The use of evolutionary optimization techniques such as genetic algorithms, differential evolution, swarm optimization and genetic programming to solve the inverse problem of parameter estimation for nonlinear chaotic systems has been gaining popularity in recent years. The efficacy of such evolutionary schemes depends on the definition of a suitable fitness function which is used to compare potential solutions in the population. In almost all research involving evolutionary schemes for parameter identification, displacement values of the first few hundred Poincaré points, after ignoring transient effects, have been used as the feature set. The measured response of the system is compared to the response of the potential solutions in the population over these Poincaré points, although there is no empirical research to show that such a feature set works better than other possible feature sets. In this paper, a smaller feature set based on first and second-order statistical parameters of the response are considered and the estimation results are compared to the estimate produced by using the standard Poincaré points-based feature set, called the finite sample feature set in this paper. Also compared are results using three evolutionary algorithms - firefly algorithm, particle swarm optimization and differential evolution. It has been shown that the proposed feature set converges to a near-optimal solution faster and in fewer generations and produces estimates that are comparable to those obtained with the finite sample feature set.
UR - http://www.scopus.com/inward/record.url?scp=84903464259&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84903464259&partnerID=8YFLogxK
U2 - 10.1115/IMECE2013-63583
DO - 10.1115/IMECE2013-63583
M3 - Conference contribution
AN - SCOPUS:84903464259
SN - 9780791856253
T3 - ASME International Mechanical Engineering Congress and Exposition, Proceedings (IMECE)
BT - Dynamics, Vibration and Control
PB - American Society of Mechanical Engineers (ASME)
T2 - ASME 2013 International Mechanical Engineering Congress and Exposition, IMECE 2013
Y2 - 15 November 2013 through 21 November 2013
ER -