A Computational Model Inspired by Gene Regulatory Networks

Abstract

Evolutionary Algorithms (EA) are parallel stochastic search procedures that are loosely inspired by the concepts of natural selection and genetic heredity. They have been successfully applied to many domains, and today Evolutionary Computation (EC) attracts a growing number of researchers from the most varied fields.

The end of the 20th century brought uncountable discoveries in the biological realm, enabled by the underlying technological breakthroughs. Complete genomes have been sequenced, including the human one, and thanks to the increasing interdisciplinarity of researchers it is known today that there is much more to evolution than just natural selection, namely the influence of the environment, gene regulation, and development. At the core of these processes there is a fundamental piece of complex biological machinery, the Genetic Regulatory Network (GRN). This network results from the interaction amongst the genes and proteins, as well as the environment, governing gene expression and consequently the development of the organism.

It is a true fact that the biological knowledge has advanced faster than our ability to incorporate it into the EAs, despite of whether or not it is benificial to do so. One of the main critics pointed-out is that the approach to the genotype-phenotype relationship is different from nature. A lot of effort has been put by some researchers into developing new representations, achieving not only improved benchmark results, but also extended flexibility and applicability of the algorithms. Moreover, others have recently started exploring computationally the new comprehension of the multitude of regulatory mechanisms that are fundamental in both the processes of inheritance and of development in natural systems, by trying to include those mechanisms in the EAs. However, few of these target machine learning problems, and most are usually developed with a specific problem domain in mind. The work presented here addresses this issue by incorporating a model of GRN in a Genetic Programming (GP) architecture.

This thesis main contribution is a model that incorporates Artificial Regulatory Networks as the genotypic representation in a GP-like system, and an algorithm that maps these networks into executable program graphs. Moreover, variants of the model were also developed, extending the capabilities of the approach to classes of problems with recursive definitions, and with multiple outputs. The efficacy and efficiency of this alternative were tested experimentally using typical benchmark problems for Genetic Programming (GP) systems, from regression to control, and logic design. Despite some limitations that were identified, the analysis of the results shows that this new method is competitive in most problem domains, even outperforming the state-of-the-art results in some cases.