[jira] [Commented] (OPENNLP-569) Fix remaining issue in L-BFGS parameter estimation to get it stable

["Vinh Khuc (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/OPENNLP-569?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13958501#comment-13958501 ] 

Vinh Khuc commented on OPENNLP-569:
-----------------------------------

Attached is the patch for the current L-BFGS implementation. It includes bug fixes for numerical overflow when calculating sum of exponential functions and the formula error when computing log-likelihood.

The patch also includes the implementation of L2-regularization to avoid overfitting.

Finally, it contains some code cleanup in QNModelReader, QNModelWriter and the unit test. 

> Fix remaining issue in L-BFGS parameter estimation to get it stable
> -------------------------------------------------------------------
>
>                 Key: OPENNLP-569
>                 URL: https://issues.apache.org/jira/browse/OPENNLP-569
>             Project: OpenNLP
>          Issue Type: Improvement
>          Components: Machine Learning
>    Affects Versions: tools-1.5.3, maxent-3.0.3
>            Reporter: Joern Kottmann
>            Assignee: Hyosup Shim
>         Attachments: LBFGS.patch
>
>
> Enhance the L-BFGS parameter estimation code to at least perform as well as the GIS training. The work on this issue should bring the implementation to a level where the experimental flag can be removed.
> For remaining problems see this issue: OPENNLP-338.



--
This message was sent by Atlassian JIRA
(v6.2#6252)