Comparison of Evolution Strategy and Back-Propagation
for Estimating Parameters of Neural Network

Roman Malczyk, Ales Gottvald

Institute of Scientific Instruments, Academy of Sciences of the CR,
Kralovopolska 147, CZ-612 64 Brno, Czech Republic

Abstract: Evolution Strategy (ES) and Back-Propagation (BP) have been compared for estimating parameters (synaptic weights and biases) of two classes of Neural Networks: Sigmoidal Neural Networks (SNN) and Green's Regularization Networks (GRN). The following features have been compared: globality of convergence, speed of convergence, asymptotic behaviour, and stability of the solutions. On average, numerical experiments show a better globality and a better asymptotic behaviour when using the (1+1)-Evolution Strategy instead of the Back-Propagation. While the speed of convergence is comparable for both the ES and BP, a simplicity of implementation makes the ES superior to BP. These conclusions have been found for both the Sigmoidal and the Green's Regularization Nets.
Back to: