Pretraining Reduces Runtime in Denoising Autoencoder Genetic Programming by an Order of Magnitude
Johannes Reiter, Dirk Schweim, David Wittenberg · 2023
Denoising autoencoder genetic programming (DAE-GP) is an estimation of distribution genetic programming (EDA-GP) algorithm. It uses denoising autoencoder long short-term memory networks as probabilistic model to replace the standard mutation and recombination operators of genetic programming (GP). Recent work has shown several advantages regarding solution length and overall performance of DAE-GP when compared to GP. However, training a neural network at each generation is computationally expensive, where model training is the most time consuming process of DAE-GP. In this work, we propose pretraining to reduce the runtime of the DAE-GP. In pretraining, the neural network is trained preceding the evolutionary search. In experiments on 8 real-world symbolic regression tasks we find that DAE-GP with pretraining has a reduced overall runtime of an order of magnitude while generating individuals with similar or better fitness.