Neuroevolution of hierarchical reservoir computers

Matthew Dale

doi:10.1145/3205455.3205520

Neuroevolution of hierarchical reservoir computers

Matthew Dale

Computer Science

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Abstract

Reservoir Computers such as Echo State Networks (ESN) represent an alternative recurrent neural network model that provides fast training and state-of-the-art performances for supervised learning problems. Classic ESNs suffer from two limitations; hyperparameter selection and learning of multiple temporal and spatial scales. To learn multiple scales, hierarchies are proposed, and to overcome manual tuning, optimisation is used. However, the autonomous design of hierarchies and optimisation of multi-reservoir systems has not yet been demonstrated. In this work, an evolvable architecture is proposed called Reservoir-of-Reservoirs (RoR) where sub-networks of neurons (ESNs) are interconnected to form the reservoir. The design of each sub-network, hyperparameters, global connectivity and its hierarchical structure are evolved using a genetic algorithm (GA) called the Microbial GA.

To evaluate the RoR and microbial GA, single networks and other hierarchical ESNs are evolved. The results show the microbial GA can dramatically increase the performance of single networks compared to other optimisation techniques. The evolutionary process also leads to competitive results with RoRs and other hierarchical ESNs, despite having fewer connections than a single network. In the final section, it is revealed that the RoR architecture may learn generalised features other architectures cannot, offering improvements in network generalisation to other tasks.

Original language	English
Title of host publication	Genetic and Evolutionary Computation Conference
Subtitle of host publication	GECCO '18: Proceedings of the Genetic and Evolutionary Computation Conference
Publisher	ACM
Pages	410-417
Number of pages	8
ISBN (Electronic)	9781450356183
DOIs	https://doi.org/10.1145/3205455.3205520
Publication status	Published - 2 Jul 2018

Access to Document

10.1145/3205455.3205520

Cite this

@inproceedings{a6a224cb0c8f435dacd843d0c2ddad22,

title = "Neuroevolution of hierarchical reservoir computers",

abstract = "Reservoir Computers such as Echo State Networks (ESN) represent an alternative recurrent neural network model that provides fast training and state-of-the-art performances for supervised learning problems. Classic ESNs suffer from two limitations; hyperparameter selection and learning of multiple temporal and spatial scales. To learn multiple scales, hierarchies are proposed, and to overcome manual tuning, optimisation is used. However, the autonomous design of hierarchies and optimisation of multi-reservoir systems has not yet been demonstrated. In this work, an evolvable architecture is proposed called Reservoir-of-Reservoirs (RoR) where sub-networks of neurons (ESNs) are interconnected to form the reservoir. The design of each sub-network, hyperparameters, global connectivity and its hierarchical structure are evolved using a genetic algorithm (GA) called the Microbial GA.To evaluate the RoR and microbial GA, single networks and other hierarchical ESNs are evolved. The results show the microbial GA can dramatically increase the performance of single networks compared to other optimisation techniques. The evolutionary process also leads to competitive results with RoRs and other hierarchical ESNs, despite having fewer connections than a single network. In the final section, it is revealed that the RoR architecture may learn generalised features other architectures cannot, offering improvements in network generalisation to other tasks.",

author = "Matthew Dale",

year = "2018",

month = jul,

day = "2",

doi = "10.1145/3205455.3205520",

language = "English",

pages = "410--417",

booktitle = "Genetic and Evolutionary Computation Conference",

publisher = "ACM",

address = "United States",

}

TY - GEN

T1 - Neuroevolution of hierarchical reservoir computers

AU - Dale, Matthew

PY - 2018/7/2

Y1 - 2018/7/2

N2 - Reservoir Computers such as Echo State Networks (ESN) represent an alternative recurrent neural network model that provides fast training and state-of-the-art performances for supervised learning problems. Classic ESNs suffer from two limitations; hyperparameter selection and learning of multiple temporal and spatial scales. To learn multiple scales, hierarchies are proposed, and to overcome manual tuning, optimisation is used. However, the autonomous design of hierarchies and optimisation of multi-reservoir systems has not yet been demonstrated. In this work, an evolvable architecture is proposed called Reservoir-of-Reservoirs (RoR) where sub-networks of neurons (ESNs) are interconnected to form the reservoir. The design of each sub-network, hyperparameters, global connectivity and its hierarchical structure are evolved using a genetic algorithm (GA) called the Microbial GA.To evaluate the RoR and microbial GA, single networks and other hierarchical ESNs are evolved. The results show the microbial GA can dramatically increase the performance of single networks compared to other optimisation techniques. The evolutionary process also leads to competitive results with RoRs and other hierarchical ESNs, despite having fewer connections than a single network. In the final section, it is revealed that the RoR architecture may learn generalised features other architectures cannot, offering improvements in network generalisation to other tasks.

AB - Reservoir Computers such as Echo State Networks (ESN) represent an alternative recurrent neural network model that provides fast training and state-of-the-art performances for supervised learning problems. Classic ESNs suffer from two limitations; hyperparameter selection and learning of multiple temporal and spatial scales. To learn multiple scales, hierarchies are proposed, and to overcome manual tuning, optimisation is used. However, the autonomous design of hierarchies and optimisation of multi-reservoir systems has not yet been demonstrated. In this work, an evolvable architecture is proposed called Reservoir-of-Reservoirs (RoR) where sub-networks of neurons (ESNs) are interconnected to form the reservoir. The design of each sub-network, hyperparameters, global connectivity and its hierarchical structure are evolved using a genetic algorithm (GA) called the Microbial GA.To evaluate the RoR and microbial GA, single networks and other hierarchical ESNs are evolved. The results show the microbial GA can dramatically increase the performance of single networks compared to other optimisation techniques. The evolutionary process also leads to competitive results with RoRs and other hierarchical ESNs, despite having fewer connections than a single network. In the final section, it is revealed that the RoR architecture may learn generalised features other architectures cannot, offering improvements in network generalisation to other tasks.

U2 - 10.1145/3205455.3205520

DO - 10.1145/3205455.3205520

M3 - Conference contribution

SP - 410

EP - 417

BT - Genetic and Evolutionary Computation Conference

PB - ACM

ER -