Universal discrete-time reservoir computers with stochastic inputs and linear readouts using non-homogeneous state-affine systems
Journal
Journal of Machine Learning Research
Type
journal article
Date Issued
2018
Author(s)
Abstract
A new class of non-homogeneous state-affine systems is introduced for use in reservoir computing. Sufficient conditions are identified that guarantee first, that the associated reservoir computers with linear readouts are causal, time-invariant, and satisfy the fading memory property and second, that a subset of this class is universal in the category of fading memory filters with stochastic almost surely uniformly bounded inputs. This means that any discrete-time filter that satisfies the fading memory property with random inputs of that type can be uniformly approximated by elements in the non-homogeneous state-affine family.
Language
English
HSG Classification
contribution to scientific community
HSG Profile Area
SEPS - Quantitative Economic Methods
Refereed
Yes
Volume
19
Start page
1
End page
40
Subject(s)
Eprints ID
258282
File(s)![Thumbnail Image]()
open.access
Name
GO_JMLR.pdf
Size
604.67 KB
Format
Adobe PDF
Checksum (MD5)
62164db87d11d2d2d43e01bacbef95b2