Deep Genetic Programming

Abstract

We propose to develop a Deep Learning (DL) framework based on the paradigm of Genetic Programming (GP). The hypothesis is that GP non-parametric and non-differentiable learning units (abstract syntax trees) have the same learning and representation capacity to Artificial Neural Networks (ANN). In an analogy to the traditional ANN/Gradient Descend/Backpropagation DL approach, the proposed framework aims at building a DL alike model fully based on GP. Preliminary results when approaching a number of application domains, suggest that GP is able to deal with large amounts of training data, such as those required in DL tasks. However, extensive research is still required regarding the construction of a multi-layered learning architecture, another hallmark of DL.