[
https://issues.apache.org/jira/browse/HAMA-809?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13790453#comment-13790453
]
Tommaso Teofili commented on HAMA-809:
--------------------------------------
I mean the one used in gradient descent, you can have a look at the test here:
https://github.com/tteofili/hama/blob/578c8f662425bf25c8aa459945931cd4b23f2972/ml/src/test/java/org/apache/hama/ml/regression/LogisticRegressionModelTest.java
and the #testMultipleCostCalculation test would fail with the current
implementation.
> Underflow in LogisticRegressionModel
> ------------------------------------
>
> Key: HAMA-809
> URL: https://issues.apache.org/jira/browse/HAMA-809
> Project: Hama
> Issue Type: Bug
> Components: machine learning
> Reporter: Tommaso Teofili
> Assignee: Tommaso Teofili
>
> There's an underflow bug in _LogisticRegressionModel_ when calculating the
> hypothesis output due to _double_ limitations.
> Therefore we should leverage _BigDecimal_s for such operations as they allow
> for more precise calculation.
> Also the log algorithm in LRM is wrong and needs to be fixed.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
