Comparative analysis of GMDH neural network based on genetic algorithm and particle swarm optimization in stable channel design.

Predicting the behavior and geometry of channels and alluvial rivers in which erosion and sediment transport are in equilibrium is among the most important topics relating to river morphology. In this study, the genetic algorithm (GA) is employed to improve the multi-objective Pareto optimal design of group method of data handling (GMDH) neural network results. The connectivity configuration in such networks is not restricted to adjacent layers. GA is applied as a new encoding scheme to generalize the structure of GMDH (GS-GMDH) for determining stable channel width based on 85 field datasets. In addition, the particle swarm optimization (PSO) learning algorithm is extended to GMDH for a better comparison of the models. The input parameters affecting channel width are the discharge, median diameter of bed sediments and Shields parameter. Sensitivity and uncertainty analyses are applied to assess the impact of each input parameter on the output parameter. The results show that GS-GMDH is more efficient than GMDH-PSO, with a high difference between predicted values. The GS-GMDH model with a correlation coefficient (R) value of 0.89 and mean absolute relative error (MARE) value of 0.053 predicted the width of a stable channel more precisely than the regression method with R of 0.81 and MARE of 0.075. The prediction uncertainty of the developed GS-GMDH indicates that the GS-GMDH model with input parameters Q, d50 and * has the least mean prediction error (MPE) of 0.019 compared with 0.116, 0.025 and 0.036 for the other models with (d50, *), (Q, d50) and (Q, *) input parameters, respectively.