Performance of Multi-Layer Perceptron Neural Networks with an Exponential Decay Activation Function in Airwaves Estimation

Activation function is one of the main characteristic elements and a vital part of Artificial Neural Networks (ANN). It is a mathematical formula present within processing unit the (neuron) of the network known to play significant role for the convergence of training algorithms. Activation function also provides nonlinear mapping potential which determine and prevent outputs of the processing unit from attaining outsized values that can inhibit training and paralyze the neural networks. Logistic (sigmoid and tangent hyperbolic) and linear are the most used activation functions in practice for the Multi-Layer Perceptron (MLP) networks. In this study, we evaluate the performance of MLP with an exponential decay function as the hidden layer activation function in the context of Marine Controlled Source Electro-Magnetic (MCSEM). Simulated airwaves data from five different sea water depths were used to evaluate the estimation performance of the MLP neural networks. The benchmark for the performance evaluation is by comparing the exponential decay activation function with log-sigmoid and tan-sigmoid activation functions through back propagation algorithm. The results seem to indicate an improved estimation performance with the exponential decay function by producing better values of Mean Square Error, Sum of Squares Error and coefficient of determination (R2).