MOSES Proposal: use SVM for program tree selection

From OpenCog
Jump to: navigation, search

Use SVM for program tree selection


Introduction

Try replacing the Bayes-net modeling in MOSES with SVM. SVM with tree kernels have now matured somewhat, see e.g.

http://www.xialuo.net/2010/11/07/svm-light-tk-1-2/

So the idea would roughly be to:

   use SVM to learn a model distinguishing fit from unfit program trees.
   use this model as a filter on generated program trees, to determine which ones are worth fitness-evaluating.

Contibutors

Currently I Jekin Trivedi am working on this project .


Timeline

Below is the expected timeline for the proposed project.

August 15 - September 13

  • Preparing for GRE exam .
  • Learning Basic of AI .

September 14 - December 15

  • Researching regarding the project .

December 16 - March 15

  • Coding starts.