A Novel Approach to Time Series Complexity via Reservoir Computing

Braden Thorne, Thomas Jüngling, Michael Small, Débora Corrêa, Ayham Zaitouny

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

When working with time series, it is often beneficial to have an idea as to how complex the signal is. Periodic, chaotic and random signals (from least to most complex) may each be approached in different ways, and knowing when a signal can be identified as belonging to one of these categories can reveal a lot about the underlying system. In the field of time series analysis, permutation entropy has emerged as one of the premier measures of time series complexity due to its ability to be calculated from data alone. We propose an alternative method for calculating complexity based on the machine learning paradigm of reservoir computing, and how the outputs of these neural networks capture similar information regarding signal complexity. We observe similar behaviour in our proposed measure to both the Lyapunov exponent and permutation entropy for well known dynamical systems. Additionally, we assess the dependence of our measure on key hyperparameters of the model, drawing conclusions about the invariance of the measure and possible implications on informing network structure.

Original languageEnglish
Title of host publicationAI 2022
Subtitle of host publicationAdvances in Artificial Intelligence - 35th Australasian Joint Conference, AI 2022, Proceedings
EditorsHaris Aziz, Débora Corrêa, Tim French
PublisherSpringer Science and Business Media Deutschland GmbH
Pages442-455
Number of pages14
ISBN (Print)9783031226946
DOIs
Publication statusPublished - 2022
Externally publishedYes
Event35th Australasian Joint Conference on Artificial Intelligence, AI 2022 - Perth, Australia
Duration: Dec 5 2022Dec 9 2022

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume13728 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference35th Australasian Joint Conference on Artificial Intelligence, AI 2022
Country/TerritoryAustralia
CityPerth
Period12/5/2212/9/22

Keywords

  • Information entropy
  • Recurrent neural networks
  • Reservoir computing
  • Time series analysis

ASJC Scopus subject areas

  • Theoretical Computer Science
  • General Computer Science

Fingerprint

Dive into the research topics of 'A Novel Approach to Time Series Complexity via Reservoir Computing'. Together they form a unique fingerprint.

Cite this