Study of neuroevolutive strategies for training and topological adaptation of artificial neural networks

• Grant number: 20/16456-7
• Support type: Scholarships in Brazil - Scientific Initiation
• Effective date (Start): May 01, 2021
• Effective date (End): April 30, 2022
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal researcher: Denis Gustavo Fantinato
• Grantee: Lucas Fernandes Muniz
• Home Institution: Centro de Matemática, Computação e Cognição (CMCC). Universidade Federal do ABC (UFABC). Ministério da Educação (Brasil). Santo André , SP, Brazil

Abstract

Artificial Neural Networks (ANN) have reached noticeable performance in many applications. In the context of supervised learning, network training is done traditionally using the backpropagation algorithm. However, this algorithm has some limitations, among which stand out its high convergence rate to local solutions and the disadvantage of assuming a fixed topology for an ANN. In this context, neuroevolution has been raised as a promising approach, in which the ANN training is done through genetic algorithms, which allow the weight adjustment alongside the topology. Especially, the algorithm NeuroEvolution of Augmenting Topologies (NEAT) has shown relevant performance in reinforcement learning tasks, due to its strategy of incremental growth from a minimum topology and the speciation mechanism. But, in the context of supervised learning, NEAT still lacks a deeper analysis. In this way, this project proposes to perform a comparative analysis between the error backpropagation algorithm and the NEAT algorithm. In this analysis, we will consider metrics from Information Theory to measure more completely the efficiency of ANNs in terms of topological complexity and network final prediction quality. Moreover, we will improve NEAT performance by combining it with the metaheuristic BRKGA and exploring its speciation mechanism. (AU)