A Study of Fitness Landscapes for Neuroevolution

Nuno M. Rodrigues, Sara Silva, Leonardo Vanneschi

Research output: Other contribution

Abstract

Fitness landscapes are a useful concept to study the dynamics of meta-heuristics. In the last two decades, they have been applied with success to estimate the optimization power of several types of evolutionary algorithms, including genetic algorithms and genetic programming. However, so far they have never been used to study the performance of machine learning algorithms on unseen data, and they have never been applied to neuroevolution. This paper aims at filling both these gaps, applying for the first time fitness landscapes to neuroevolution and using them to infer useful information about the predictive ability of the method. More specifically, we use a grammar-based approach to generate convolutional neural networks, and we study the dynamics of three different mutations to evolve them. To characterize fitness landscapes, we study autocorrelation and entropic measure of ruggedness. The results show that these measures are appropriate for estimating both the optimization power and the generalization ability of the considered neuroevolution configurations.
Original languageEnglish
TypeE-print
PublisherCornell University (ArXiv)
Number of pages8
Publication statusPublished - 30 Jan 2020

Publication series

NamearXiv.org
No.2001.11272
VolumeComputer Science. Neural and Evolutionary Computing

Fingerprint

Genetic programming
Autocorrelation
Evolutionary algorithms
Learning algorithms
Learning systems
Genetic algorithms
Neural networks

Keywords

  • Fitness Landscapes
  • Entropic Measure of Ruggedness
  • Neuroevolution
  • Convolutional Neural Networks
  • Autocorrelation

Cite this

Rodrigues, N. M., Silva, S., & Vanneschi, L. (2020, Jan 30). A Study of Fitness Landscapes for Neuroevolution. Cornell University (ArXiv).
Rodrigues, Nuno M. ; Silva, Sara ; Vanneschi, Leonardo. / A Study of Fitness Landscapes for Neuroevolution. 2020. Cornell University (ArXiv). 8 p. (arXiv.org; 2001.11272).
@misc{07a39b44f21b4816a3605a5ee17f6e5b,
title = "A Study of Fitness Landscapes for Neuroevolution",
abstract = "Fitness landscapes are a useful concept to study the dynamics of meta-heuristics. In the last two decades, they have been applied with success to estimate the optimization power of several types of evolutionary algorithms, including genetic algorithms and genetic programming. However, so far they have never been used to study the performance of machine learning algorithms on unseen data, and they have never been applied to neuroevolution. This paper aims at filling both these gaps, applying for the first time fitness landscapes to neuroevolution and using them to infer useful information about the predictive ability of the method. More specifically, we use a grammar-based approach to generate convolutional neural networks, and we study the dynamics of three different mutations to evolve them. To characterize fitness landscapes, we study autocorrelation and entropic measure of ruggedness. The results show that these measures are appropriate for estimating both the optimization power and the generalization ability of the considered neuroevolution configurations.",
keywords = "Fitness Landscapes, Entropic Measure of Ruggedness, Neuroevolution, Convolutional Neural Networks, Autocorrelation",
author = "Rodrigues, {Nuno M.} and Sara Silva and Leonardo Vanneschi",
note = "Rodrigues, N. M., Silva, S., & Vanneschi, L. (2020, Jan 30). A Study of Fitness Landscapes for Neuroevolution. (arXiv.org; Vol. Computer Science. Neural and Evolutionary Computing, No. 2001.11272). Cornell University (ArXiv).",
year = "2020",
month = "1",
day = "30",
language = "English",
series = "arXiv.org",
publisher = "Cornell University (ArXiv)",
number = "2001.11272",
type = "Other",

}

Rodrigues, NM, Silva, S & Vanneschi, L 2020, A Study of Fitness Landscapes for Neuroevolution. Cornell University (ArXiv).

A Study of Fitness Landscapes for Neuroevolution. / Rodrigues, Nuno M. ; Silva, Sara; Vanneschi, Leonardo.

8 p. Cornell University (ArXiv). 2020, E-print. (arXiv.org; Vol. Computer Science. Neural and Evolutionary Computing, No. 2001.11272).

Research output: Other contribution

TY - GEN

T1 - A Study of Fitness Landscapes for Neuroevolution

AU - Rodrigues, Nuno M.

AU - Silva, Sara

AU - Vanneschi, Leonardo

N1 - Rodrigues, N. M., Silva, S., & Vanneschi, L. (2020, Jan 30). A Study of Fitness Landscapes for Neuroevolution. (arXiv.org; Vol. Computer Science. Neural and Evolutionary Computing, No. 2001.11272). Cornell University (ArXiv).

PY - 2020/1/30

Y1 - 2020/1/30

N2 - Fitness landscapes are a useful concept to study the dynamics of meta-heuristics. In the last two decades, they have been applied with success to estimate the optimization power of several types of evolutionary algorithms, including genetic algorithms and genetic programming. However, so far they have never been used to study the performance of machine learning algorithms on unseen data, and they have never been applied to neuroevolution. This paper aims at filling both these gaps, applying for the first time fitness landscapes to neuroevolution and using them to infer useful information about the predictive ability of the method. More specifically, we use a grammar-based approach to generate convolutional neural networks, and we study the dynamics of three different mutations to evolve them. To characterize fitness landscapes, we study autocorrelation and entropic measure of ruggedness. The results show that these measures are appropriate for estimating both the optimization power and the generalization ability of the considered neuroevolution configurations.

AB - Fitness landscapes are a useful concept to study the dynamics of meta-heuristics. In the last two decades, they have been applied with success to estimate the optimization power of several types of evolutionary algorithms, including genetic algorithms and genetic programming. However, so far they have never been used to study the performance of machine learning algorithms on unseen data, and they have never been applied to neuroevolution. This paper aims at filling both these gaps, applying for the first time fitness landscapes to neuroevolution and using them to infer useful information about the predictive ability of the method. More specifically, we use a grammar-based approach to generate convolutional neural networks, and we study the dynamics of three different mutations to evolve them. To characterize fitness landscapes, we study autocorrelation and entropic measure of ruggedness. The results show that these measures are appropriate for estimating both the optimization power and the generalization ability of the considered neuroevolution configurations.

KW - Fitness Landscapes

KW - Entropic Measure of Ruggedness

KW - Neuroevolution

KW - Convolutional Neural Networks

KW - Autocorrelation

M3 - Other contribution

T3 - arXiv.org

PB - Cornell University (ArXiv)

ER -