Using Denoising Autoencoder Genetic Programming to Control Exploration and Exploitation in Search

Abstract

Denoising Autoencoder Genetic Programming (DAE-GP) is a novel neural network-based estimation of distribution genetic programming (EDA-GP) algorithm that uses denoising autoencoder long short-term memory networks as a probabilistic model to replace the standard mutation and recombination operators of genetic programming (GP). At each generation, the idea is to flexibly identify promising properties of the parent population and to transfer these properties to the offspring where the DAE-GP uses denoising to make the model robust to noise that is present in the parent population. Denoising partially corrupts candidate solutions that are used as input to the model. The stronger the corruption, the stronger the generalization of the model. In this work, we study how corruption strength affects the exploration and exploitation behavior of the DAE-GP. For a generalization of the royal tree problem (high-locality problem), we find that the stronger the corruption, the stronger the exploration of the solution space. For the given problem, weak corruption resulting in a stronger exploitation of the solution space performs best. However, in more rugged fitness landscapes (low-locality problems), we expect that a stronger corruption resulting in a stronger exploration will be helpful. Choosing the right denoising strategy can therefore help to control the exploration and exploitation behavior in search, leading to an improved search quality.