Document Type

Conference Paper

Department/Unit

Department of Computer Science

Title

On compressibility and acceleration of orthogonal NMF for POMDP compression

Language

English

Abstract

State space compression is one of the recently proposed approaches for improving POMDP's tractability. Despite its initial success, it still carries two intrinsic limitations. First, not all POMDP problems can be compressed equally well. Also, the cost of computing the compressed space itself may become significant as the size of the problem is scaled up. In this paper, we address the two issues with respect to an orthogonal non-negative matrix factorization recently proposed for POMDP compression. In particular, we first propose an eigenvalue analysis to evaluate the compressibility of a POMDP and determine an effective range for the dimension reduction. Also, we incorporate the interior-point gradient acceleration into the orthogonal NMF and derive an accelerated version to minimize the compression overhead. The validity of the eigenvalue analysis has been evaluated empirically. Also, the proposed accelerated orthogonal NMF has been demonstrated to be effective in speeding up the policy computation for a set of robot navigation related problems.

Keywords

Optimal Policy, Belief State, Krylov Subspace, Robot Navigation, Eigenvalue Analysis

Publication Date

11-2009

Source Publication Title

Advances in Machine Learning: First Asian Conference on Machine Learning, ACML 2009, Nanjing, China, November 2-4, 2009. Proceedings

Editors

Zhou, Zhi-Hua ; Washio, Takashi

Start Page

248

End Page

262

Series Title

Lecture notes in computer science, 5828.

Conference Location

Nanjing, China

Publisher

Springer

Peer Reviewed

1

Funder

This work is partially supported by HKBU Faculty Research Grant (FRG/08- 09/124).

DOI

10.1007/978-3-642-05224-8_20

Link to Publisher's Edition

http://dx.doi.org/10.1007/978-3-642-05224-8_20

ISBN (print)

9783642052231

ISBN (electronic)

9783642052248

This document is currently not available here.

Share

COinS