Nonlinear time series prediction using modified BBO-based trained TDCMAC network.

Nonlinear time series analysis and forecasting are an essential part in a diverse range of physical and natural applications. This paper presents a time-division cerebellar model articulation controller (TDCMAC) network using a modified biogeography-based optimization (modified BBO) learning algorithm for nonlinear time series and measurement data prediction. The TDCMAC method is a time windowing strategy constructed using the CMAC network. BBO algorithm is designed related to the geographical distribution of species over time and space. This study presents two modified migration functions of the essential BBO, i.e. quadratic migration BBO (QBBO) and sinusoidal migration BBO (SBBO) methods, to improve convergence rate and quality of solution. Five nonlinear time series, including Mackey-glass, Lorenz, Rossler, Limber Pine, and Ponderosa Pine data series, are employed to investigate the proposed predictor. The TDCMAC networks using QBBO and SBBO learning algorithms are compared with the gradient descent (GD) method and other existing heuristic learning methods, including particle swarm optimization (PSO), genetic algorithm (GA), and conventional BBO methods, to verify the estimation performance of the proposed method. The performances are evaluated through an extensive simulation by computing the root mean square error (RMSE), mean absolute percentage error (MAPE), and average relative variance (ARV) metrics. Experimental results demonstrate that the proposed predictor indeed achieve more accurate performances and faster learning speed for time series prediction applications. [ABSTRACT FROM AUTHOR]