multisearch-weka-package

Weka package for parameter optimization, similar to GridSearch. Can be used for optimizing an arbitrary number of parameters, in contrast to GridSearch which always requires you to optimize two parameters. However, it does not offer automatic search space extensions like GridSearch.

Parameters

• weka.core.setupgenerator.MathParameter

  uses mathematical expression to compute the (numeric) parameters

• weka.core.setupgenerator.ListParameter

  list of values; default is blank-separated, but a custom delimiter can be supplied

• weka.core.setupgenerator.MLPLayersParameter

  generates hidden layer definitions for the MultiLayerPerceptron classifier (contributed by Jan van Rijn)

• weka.core.setupgenerator.ParameterGroup

  allows grouping of dependent parameters, e.g., setting on group sets the kernel of SMO to RFBKernel and explores the gamma option, another group sets the kernel to PolyKernel and explores the exponent option.

Note: array elements, e.g., the filters inside a weka.filters.MultiFilter can be accessed using [n] with n being the 0-based index. E.g., if the third filter inside a MultiFilter is a PLSFilter, then its numComponents property can be accessed with filters[2].numComponents.

Supported parameter types

• char, string
• float, double
• int, long
• boolean
• weka.core.SelectedTag
• Java classname (and possible options for classes implementing weka.core.OptionHandler)

Search space exploration

The search space of setups can be explored with different strategies, derived from weka.classifiers.meta.multisearch.AbstractSearch. The following strategies are available:

• weka.classifiers.meta.multisearch.DefaultSearch

  Exhaustive search of parameter space

• weka.classifiers.meta.multisearch.RandomSearch

  Random search of parameter space (contributed by Jan van Rijn)

Example

With the following classifier setup:

weka.classifiers.meta.FilteredClassifier
|
+- weka.filters.supervised.attribute.PLSFilter
|
+- weka.classifiers.functions.LinearRegression

You can explore the filter's PLS components and classifier's ridge parameters by referencing these properties as follows (the MultiSearch's classifier is used as the base for the property paths):

• components: filter.numComponents
• ridge: classifier.ridge

For more examples, please see the following repository:

https://github.com/fracpete/multisearch-weka-package-examples

Releases

Click on one of the following links to download the corresponding Weka package:

• 2021.2.17
• 2020.2.17
• 2019.10.4
• 2019.9.30
• old releases: 2018.9.2, 2018.8.16, 2018.8.11, 2017.10.1, 2017.3.28, 2016.6.6, 2016.5.31, 2016.5.25, 2016.5.15, 2016.4.30, 2016.4.25, 2016.4.14, 2016.1.30, 2016.1.15, 2016.1.14, 2016.1.13, 2015.10.15, 2015.9.2, 2014.12.10

How to use packages

For more information on how to install the package, see:

https://waikato.github.io/weka-wiki/packages/manager/

Maven

Add the following dependency in your pom.xml to include the package:

    <dependency>
      <groupId>com.github.fracpete</groupId>
      <artifactId>multisearch-weka-package</artifactId>
      <version>2021.2.17</version>
      <type>jar</type>
      <exclusions>
        <exclusion>
          <groupId>nz.ac.waikato.cms.weka</groupId>
          <artifactId>weka-dev</artifactId>
        </exclusion>
      </exclusions>
    </dependency>