Back to search

University of York

Breaking Implicit Assumptions of Physical Delay-Feedback Reservoir Computing

Abstract

dc:description.abstract

The Reservoir Computing (RC) paradigm is a supervised machine learning scheme using the natural computational ability of dynamical systems. Such dynamical systems incorporate time delays showcasing intricate dynamics. This richness in dynamics, particularly the system's transient response to external stimuli makes them suitable for RC. A subset of RCs, Delay-Feedback Reservoir Computing (DFRC), is distinguished by its unique features: a system that consists of a single nonlinear node and a delay-line, with `virtual' nodes defined along the delay-line by time-multiplexing procedure of the input. These characteristics render DFRC particularly useful for hardware integration. In this thesis, the aim is to break the implicit assumptions made in the design of physical DFRC based on Mackey-Glass dynamical system. The first assumption we address is the performance of DFRC is not affected by the attenuation in physcial delay-line as the nodes defined along it are 'virtual'. However, our experimental results contradict this. To mitigate the impact of losses along the delay line, we propose a methodology `Devirtualisation', which describes the procedure of directly tapping into the delay lines at the position of a `virtual' node, rather than at the delay line's end. It trade-offs the DFRC system's read-out frequency and the quantity of output lines. Masking plays a crucial role in DFRC, as it defines `virtual' nodes along the delay-line. The second assumption is that the mask used should randomly generated numbers uniformly distributed between [-u,u]. We experimentally compare Binary Weight Mask (BWM) vs. Random Weight Mask (RWM) under different scenarios; and investigate the randomness of BWM signal distribution's impact. The third implicit assumption is that, DFRC is designed to solve time series prediction tasks involving a single input and output with no external feedback. To break this assumption, we propose two approaches to mix multi-input signals into DFRC; to validate these approaches, a novel task for DFRC that inherently necessitates multiple inputs: the control of a forced Van der Pol oscillator system, is proposed.

Degree

thesis:*
Name dc:type.qualificationname
Ph.D
Level dc:type.qualificationlevel
doctoral
Grantor dc:publisher.institution
University of York
Year dc:date.issued
2023

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Gan, Tian
Advisors dc:contributor.advisor
  • Trefzer, Martin
  • Stepney, Susan
ORCID person:identifier.orcid

Rights

Language dc:language.dcterms_iso639_2
eng

Chain of custody

source
Harvested from
White Rose University Consortium
Base URL
etheses.whiterose.ac.uk/cgi/oai2
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
related terms
citation

Gan, Tian. Breaking Implicit Assumptions of Physical Delay-Feedback Reservoir Computing. doctoral thesis, University of York, 2023.