Minimizing the EXA-GP Graph-Based Genetic Programming Algorithm for Interpretable Time Series Forecasting

This work provides a modification to the Evolutionary eXploration of Augmenting Memory Models (EXA-GP) graph-based genetic programming (GGP) algorithm, enabling it to produce time series forecasting (TSF) equations that are vastly more simple and interpretable than the original implementation without heavily compromising on predictive ability. This is accomplished by eliminating a majority of all the trainable constants and initializing the algorithm with a seed computational graph in the form of using a parameter's value at time t as the forecast for the parameter's value at time t + 1. This minimal version of EXA-GP (EXA-GP-MIN) is compared to EXA-GP and EXAMM, a full blown neuroevolution algorithm for evolving recurrent neural networks for TSF, on a suite of six real world benchmark problems, with MIN-EXA-GP showing the best forecasting ability on four of the six benchmarks with significantly more interpretable genetic programs.