Paper Abstract and Keywords |
Presentation |
2011-06-24 16:30
Solving POMDPs using Restricted Boltzmann Machines with Echo State Networks Makoto Otsuka, Junichiro Yoshimoto, Stefan Elfwing, Kenji Doya (OIST) NC2011-19 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
A partially observable Markov decision process (POMDP) can be solved in a model-based way using explicit knowledge of the environmental dynamics or in a model-free way using implicit representations of task-relevant states. Here we consider a model-free approach of combining an echo state network (ESN) for summarizing past actions and observations and a restricted Boltzmann machine (RBM) for learning action values in a high-dimensional state space. Simulation results in robot navigation tasks showed that the ESN can capture relevant information in the sequence of high dimensional observations and that RBM can construct task-oriented internal representation in its hidden layer. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
partially observable Markov decision processes / restricted Boltzmann machines / echo state networks / goal-directed representation / free energy / / / |
Reference Info. |
IEICE Tech. Rep., vol. 111, no. 96, NC2011-19, pp. 143-148, June 2011. |
Paper # |
NC2011-19 |
Date of Issue |
2011-06-16 (NC) |
ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
Copyright and reproduction |
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
Download PDF |
NC2011-19 |