Symbolic Regression using Genetic Programming Leveraging Neural Information Processing

Abstract: Regression analysis conducted with traditional mathematical methods can be sub-optimal if the exact model of the observed data is unknown. Evolutionary computing (EC) and deep learning (DL) are viable alternatives, since regression performed with these methods tends to be less dependent on a particular model. EC are especially flexible, because they are capable of performing symbolic regression. A subfield of EC and DL is genetic programming (GP) and artificial neural networks (ANN), respectively. This master thesis examines the effects of giving a genetic programming system neural information processing capabilities, in order to bridge the gap between ANN and GP. The approach is to compare GP, in its standard formulation, with 1)~GP that speciates using an ANN, 2)~GP that extends the function set with ANNs. Two methods are used to measure the prediction error. The effect of the first approach is an increased noise in the convergence. This leads to an enlarged spread of the prediction error for one of our two error measures, and a mainly unchanged error for the other. The effects of the second approach is an increase in accuracy for one of the error measures, and a decrease in bloat.