• English
    • français
    • Deutsch
    • español
    • português (Brasil)
    • Bahasa Indonesia
    • русский
    • العربية
    • 中文
  • English 
    • English
    • français
    • Deutsch
    • español
    • português (Brasil)
    • Bahasa Indonesia
    • русский
    • العربية
    • 中文
  • Login
View Item 
  •   Home
  • OAI Data Pool
  • OAI Harvested Content
  • View Item
  •   Home
  • OAI Data Pool
  • OAI Harvested Content
  • View Item
JavaScript is disabled for your browser. Some features of this site may not work without it.

Browse

All of the LibraryCommunitiesPublication DateTitlesSubjectsAuthorsThis CollectionPublication DateTitlesSubjectsAuthorsProfilesView

My Account

LoginRegister

The Library

AboutNew SubmissionSubmission GuideSearch GuideRepository PolicyContact

Compressed predictive state representation: an efficient moment-method for sequence prediction and sequential decision-making

  • CSV
  • RefMan
  • EndNote
  • BibTex
  • RefWorks
Author(s)
Hamilton, William
Contributor(s)
Joelle Pineau (Internal/Supervisor)
Keywords
Applied Sciences - Computer Science

Full record
Show full item record
URI
http://hdl.handle.net/20.500.12424/842610
Online Access
http://digitool.Library.McGill.CA:80/R/?func=dbin-jump-full&object_id=127091
Abstract
The construction of accurate predictive models over sequence data is of fundamental importance in the pursuit of developing artificial agents that are capable of (near)-optimal sequential decision-making in disparate environments.If such predictive models are available, they can be exploited by decision-making agents, allowing them to reason about the future dynamics of a system. Constructing models with sufficient predictive capacity, however, is a difficult task. Under the standard maximum likelihood criterion, these models tend to have non-convex learning objectives, and heuristics such as expectation-maximization suffer from high computational overhead. In contrast, an alternative statistical objective, the so-called method of moments, leads to convex optimizations that are often efficiently solvable via spectral decompositions.This work further improves upon the scalability, efficiency, and accuracy of this moment-based framework by employing techniques from the field of compressed sensing. Specifically, random projections of high-dimensional data are used during learning to (1) provide computational efficiency and (2) regularize the learned predictive models. Both theoretical analyses, outlining an explicit bias-variance trade-off, and experiments, demonstrating the superior empirical performance of the novel algorithm (e.g., compared to uncompressed moment-methods), are provided. Going further, this work introduces a sequential decision-making framework which exploits these compressed learned models. Experiments demonstrate that the combination of the compressed model learning algorithm and this decision-making framework allows for agents to successfully plan in massive, complex environments without prior knowledge.
Pour pouvoir agir optimalement, il est important de pouvoir prédire les séquences d'observations a venir. Ceci est une tache difficile puisque le problème d'estimation du maximum de vraisemblance est non-convexe. Les méthodes standard, tel que l'algorithme d'espérance-maximisation, sont très coûteuses et inefficaces. Une application récente de la méthode des moments offre une interprétation différente du problème qui est convexe et efficace.La méthode présentée améliore l'efficacité et la précision de la méthode des moments dans le contexte de prédiction de séquences d'observations. Ceci est fait grâce à des projections aléatoires qui augmentent l'efficacité de l'algorithme. Une analyse théorique de notre méthode démontre que notre algorithme réduit la variance au dépend d'un peu plus de biais. Nos résultats empiriques démontret une meilleure performance comparativement aux méthodes précédentes. De plus, nous offrons un moyen d'exploiter les prédictions de notre algorithme de façon à agir optimalement. Ceci nous permet de produire des agents capable de raisonner dans des environnements complexes et partiellement observables.
Date
2014
Type
Electronic Thesis or Dissertation
Identifier
oai:digitool.library.mcgill.ca:127091
http://digitool.Library.McGill.CA:80/R/?func=dbin-jump-full&object_id=127091
Copyright/License
All items in eScholarship@McGill are protected by copyright with all rights reserved unless otherwise indicated.
Collections
OAI Harvested Content

entitlement

 
DSpace software (copyright © 2002 - 2021)  DuraSpace
Quick Guide | Contact Us
Open Repository is a service operated by 
Atmire NV
 

Export search results

The export option will allow you to export the current search results of the entered query to a file. Different formats are available for download. To export the items, click on the button corresponding with the preferred download format.

By default, clicking on the export buttons will result in a download of the allowed maximum amount of items.

To select a subset of the search results, click "Selective Export" button and make a selection of the items you want to export. The amount of items that can be exported at once is similarly restricted as the full export.

After making a selection, click one of the export format buttons. The amount of items that will be exported is indicated in the bubble next to export format.