Applying optimized BPN to a chaotic time series problem

Back-propagation network (BPN), the most often used ANN, has been employed to solve a number of real problems. It has a tendency to be trapped at local minima, i.e. non-global optimal solutions as applying to the forecasting problems according to the past experience. These limitations cause the BPN to become inconsistent and unpredictable. On the other hand, because the oscillations and irregular motions have often been observed in real time series, deterministic equilibrium models could not describe the phenomena of real data in forecasting modeling. Consequently, the nonlinear chaotic model has attracted researcher's attention as a possible alternative explanation for the fluctuations phenomena in the real world because of its ability to offer a way to produce mentioned above behavior without the introduction of stochastic elements. This study will use the genetic algorithm to optimize BPN in order to forecast the chaotic time series problem truthfully. The experimental results confirm that the genetic algorithm performs well as a global search algorithm. Furthermore, it is shown that designating the topology and parameters of the neural network as decision variables simultaneously and then using the genetic algorithms to determine their values can improve the forecasting effectiveness of the resulting BPN when applied to a chaotic time series problem.