Constructive ANN with Dynamically Set Sigmoid: A Simulation Tool for Technoeconomic Forecasting

Abstract

To forecast future outcomes in an economic growth scenario, a novel and compatible version of an artificial neural network (ANN) is developed. It is designed with a dynamic sigmoidal (squashing) function that morphs to the stochastical trends of the ANN input. Relevant network architecture then gets pruned for reduced complexity across the span of iterative training schedule leading to the realisation of a constructive artificial neural-network (CANN). In the associated operation, the input-output relation in the network is set to change dynamically by varying the initial slope of the intervening squashing function (sigmoid) during the training phase. This slope-changing is realised by choosing the so-called Langevin-Bernoulli function (LBF) as the sigmoid, (in lieu of traditional hyperbolic tangent function). By changing the associated order-parameter, the initial slope of LBF is altered enabling an eventual reduced network complexity. A CANN designed with above features is used to track the temporal evolution of an entity (such as an economic parameter); and, it is adopted to forecast future trends of technoeconomic evolution with minimal under- or over-fitted predictions. The conceived method is applied to an available set of technoeconomic data and the efficacy of forecasting is ascertained thereof.