Change search
ReferencesLink to record
Permanent link

Direct link
A Novel Method for Training an Echo State Network with Feedback-Error Learning
Norwegian University of Science and Technology, Faculty of Information Technology, Mathematics and Electrical Engineering, Department of Computer and Information Science.
2013 (English)In: Advances in Artificial Intelligence, ISSN 1687-7470, E-ISSN 1687-7489, Vol. 2013Article in journal (Refereed) Published
Abstract [en]

Echo state networks are a relatively new type of recurrent neural networks that have shown great potentials for solving non-linear, temporal problems. The basic idea is to transform the low dimensional temporal input into a higher dimensional state, and then train the output connection weights to make the system output the target information. Because only the output weights are altered, training is typically quick and computationally efficient compared to training of other recurrent neural networks. This paper investigates using an echo state network to learn the inverse kinematics model of a robot simulator with feedback-error-learning. In this scheme teacher forcing is not perfect, and joint constraints on the simulator makes the feedback error inaccurate. A novel training method which is less influenced by the noise in the training data is proposed and compared to the traditional ESN training method.

Place, publisher, year, edition, pages
2013. Vol. 2013
URN: urn:nbn:no:ntnu:diva-23190DOI: 10.1155/2013/891501OAI: diva2:658937

Copyright © 2013 Rikke Amilde Lovlid. This is an open access article distributed under theCreativeCommonsAttribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Available from: 2013-10-23 Created: 2013-10-23 Last updated: 2013-11-25Bibliographically approved
In thesis
1. Internal Models as Echo State Networks: Learning to Execute Arm Movements
Open this publication in new window or tab >>Internal Models as Echo State Networks: Learning to Execute Arm Movements
2013 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

As robots are becoming more and more complex, with higher degrees-of-freedom, lighter limbs, and springy joints, it becomes harder to control their movements. New approaches, inspired from neuroscience, are attracting increased attention among computer scientists dealing with motor control.

The focus in this thesis is on how robots can learn to control their limbs by learning how their bodies work, i.e., by learning internal models of their motor apparatus. Inspiration from cerebellar research combined with concepts from traditional control theory has been used as a basis.

The research in the thesis is twofold. First, we investigate how internal models can be used to solve different control problems. In particular, we consider how to handle delays in the sensory-motor-loop and how to realize bimanual coordination. Second, we study how the internal models can be represented and learned. This includes how to choose movements to learn from in order to learn as much of the internal model as possible and how to actually learn the training movement.

A simple simulator is used in the experiments and the simulator’s internal models were implemented as echo state networks (ESNs), a relatively new and promising type of recurrent neural networks. The simulator learns internal modes of his motor apparatus by imitating human motion. Human motion data was recorded and the task of the simulator’s control system is to generate motor commands that result in the simulator replicating the recorded movement.

From the experiments we conclude that using ESNs for representing and learning internal models looks promising. With an ESN we are able to generalize to imitating novel movements, and we demonstrate that it is able to learn various bimanual coordination patterns. However, training ESNs is challenging and a major contribution from this thesis is a novel training method that works particularly well in our application. The thesis also contributes to how different internal models can be used and trained together.

Place, publisher, year, edition, pages
NTNU: , 2013
Doctoral theses at NTNU, ISSN 1503-8181 ; 2013:335
National Category
Computer technology Information and communication science
urn:nbn:no:ntnu:diva-23554 (URN)978-82-471-4810-5 (printed ver.) (ISBN)978-82-471-4811-2 (electronic ver.) (ISBN)
Public defence
2013-12-09, 00:00
Available from: 2013-12-09 Created: 2013-11-25 Last updated: 2014-01-08Bibliographically approved

Open Access in DiVA

fulltext(2725 kB)35 downloads
File information
File name FULLTEXT01.pdfFile size 2725 kBChecksum SHA-512
Type fulltextMimetype application/pdf

Other links

Publisher's full textAdvances in Artificial Intelligence
By organisation
Department of Computer and Information Science
In the same journal
Advances in Artificial Intelligence

Search outside of DiVA

GoogleGoogle Scholar
Total: 35 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

Altmetric score

Total: 26 hits
ReferencesLink to record
Permanent link

Direct link