Reading PAGE

Peer Evaluation activity

Reviews 1
Downloads 2
Views 330
Followed by 1

Total impact ?

    Send a

    Tom has...

    Trusted 0
    Reviewed 0
    Emailed 0
    Shared/re-used 0
    Discussed 0
    Invited 0
    Collected 0

     

    This was brought to you by:

    block this user Tom Schaul

    Post Doctorate

    Courant Institute, New York University

    Episodic Reinforcement Learning by Logistic Reward-Weighted Regression

    Export to Mendeley

    It has been a long-standing goal in the adaptive control community to reduce the generically difficult, general reinforcement learning (RL) problem to simpler problems solvable by supervised learning. While this approach is todays standard for value function-based methods, fewer approaches are known that apply similar reductions to policy search methods. Recently, it has been shown that immediate RL problems can be solved by reward-weighted regression, and that the resulting algorithm is an expectation maximization (EM) algorithm with strong guarantees. In this paper, we extend this algorithm to the episodic case and show that it can be used in the context of LSTM recurrent neural networks (RNNs). The resulting RNN training algorithm is equivalent to a weighted self-modeling supervised learning technique. We focus on partially observable Markov decision problems (POMDPs) where it is essential that the policy is nonstationary in order to be optimal. We show that this new reward-weighted logistic regression u sed in conjunction with an RNN architecture can solve standard benchmark POMDPs with ease.

    Oh la laClose

    Your session has expired but don’t worry, your message
    has been saved.Please log in and we’ll bring you back
    to this page. You’ll just need to click “Send”.

    Your evaluation is of great value to our authors and readers. Many thanks for your time.

    Review Close

    Short review
    Select a comment
    Select a grade
    You and the author
    Anonymity My review is anonymous( Log in  or  Register )
    publish
    Close

    When you're done, click "publish"

    Only blue fields are mandatory.

    Relation to the author*
    Overall Comment*
    Anonymity* My review is anonymous( Log in  or  Register )
     

    Focus & Objectives*

    Have the objectives and the central topic been clearly introduced?

    Novelty & Originality*

    Do you consider this work to be an interesting contribution to knowledge?

    Arrangement, Transition and Logic

    Are the different sections of this work well arranged and distributed?

    Methodology & Results

    Is the author's methodology relevant to both the objectives and the results?

    Data Settings & Figures

    Were tables and figures appropriate and well conceived?

    References and bibliography

    Is this work well documented and has the bibliography been properly established?

    Writing

    Is this work well written, checked and edited?

    Write Your Review (you can paste text as well)
    Please be civil and constructive. Thank you.


    Grade (optional, N/A by default)

    N/A 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10
    Close

    Your mailing list is currently empty.
    It will build up as you send messages
    and links to your peers.

     No one besides you has access to this list.
    Close
    Enter the e-mail addresses of your recipients in the box below.  Note: Peer Evaluation will NOT store these email addresses   log in
    Your recipients

    Your message:

    Your email : Your email address will not be stored or shared with others.

    Your message has been sent.

    Description

    Title : Episodic Reinforcement Learning by Logistic Reward-Weighted Regression
    Author(s) : Daan Wierstra, Tom Schaul, Jan Peters, Juergen Schmidhuber
    Abstract : It has been a long-standing goal in the adaptive control community to reduce the generically difficult, general reinforcement learning (RL) problem to simpler problems solvable by supervised learning. While this approach is todays standard for value function-based methods, fewer approaches are known that apply similar reductions to policy search methods. Recently, it has been shown that immediate RL problems can be solved by reward-weighted regression, and that the resulting algorithm is an expectation maximization (EM) algorithm with strong guarantees. In this paper, we extend this algorithm to the episodic case and show that it can be used in the context of LSTM recurrent neural networks (RNNs). The resulting RNN training algorithm is equivalent to a weighted self-modeling supervised learning technique. We focus on partially observable Markov decision problems (POMDPs) where it is essential that the policy is nonstationary in order to be optimal. We show that this new reward-weighted logistic regression u sed in conjunction with an RNN architecture can solve standard benchmark POMDPs with ease.
    Subject : unspecified
    Area : Other
    Year : 2008

    Affiliations Courant Institute, New York University
    Editors : Véra Kůrková, Roman Neruda, Jan Koutník
    Journal : Artificial Neural Networks ICANN 2008
    Volume : 5163
    Issue : 1
    Publisher : Springer Berlin Heidelberg
    City : Berlin, Heidelberg
    Pages : 407 - 416
    Url : http://www.springerlink.com/index/10.1007/978-3-540-87536-9_42
    Isbn : 978-3-540-87535-2
    Doi : 10.1007/978-3-540-87536-9_42

    Leave a comment

    This contribution has not been reviewed yet. review?

    You may receive the Trusted member label after :

    • Reviewing 10 uploads, whatever the media type.
    • Being trusted by 10 peers.
    • If you are blocked by 10 peers the "Trust label" will be suspended from your page. We encourage you to contact the administrator to contest the suspension.

    Does this seem fair to you? Please make your suggestions.

    Please select an affiliation to sign your evaluation:

    Cancel Evaluation Save

    Please select an affiliation:

    Cancel   Save

    Tom's Peer Evaluation activity

    Tom has...

    Trusted 0
    Reviewed 0
    Emailed 0
    Shared/re-used 0
    Discussed 0
    Invited 0
    Collected 0
    Invite this peer to...
    Title
    Start date (dd/mm/aaaa)
    Location
    URL
    Message
    send
    Close

    Full Text request

    Your request will be sent.

    Please enter your email address to be notified
    when this article becomes available

    Your email


     
    Your email address will not be shared or spammed.