Reply to the discussion by Das and Sivakugan on “Intelligent computing for modeling axial capacity of pile foundations”Appears in Canadian Geotechnical Journal, 47(8): 928–930.

Abstract

The author welcomes and appreciates the interest and comments of the discussers, and the issues that they have raised are responded to as follows. The discussers recommend scaling the inputs and outputs data between [–1, 1] rather than [0, 1] because the transfer function used in the hidden layer was tanh, which is bounded by [–1, 1]. It should be noted that the sigmoid transfer function was used in the output layer and thus, the outputs must be scaled between [0, 1], commensurate with the limits of the sigmoid transfer function. Using the tanh transfer function in the output layer is not acceptable, as it does not comply with the underlying physical meaning of the bearing capacity problem. To explain this, consider the minimum value of ultimate bearing capacity of 290 kN and the maximum value of 4500 kN that was used in the driven piles model. If these two values are to be scaled between [–1, 1], the minimum ultimate bearing capacity will be equal to –1 and the maximum ultimate bearing capacity will be equal to 1. This means that, during artificial neural network (ANN) model training, the impact of the minimum ultimate bearing capacity will be as if it is exactly opposite to the impact of the maximum ultimate bearing capacity, which is not true from the physical point of view. In the hidden layer, however, the tanh transfer function was used, as it was found to