I have one question regarding the Machine learning (ML) /Evolutionary Algorithm(EA) method. One of my research objectives is to develop new equation taking as a starting point the pool of possible variables and the pool of simple arithmetic operators; the new equation resulting from the ML/EA method would then be evaluated by the extent to which they could correctly classify the training-set.
Could you suggest, which ML / EA method should I use to deliver the objective? Cartesian Genetic Programming or Basic GP or Neural Network or Deep learning or etc?
Your help in this matter would be greatly appreciated.
.
if i understand well, your approach to feature engineering is to have your variables and operators sets and combine them in some way, obtain a set of features and then evaluate this set for the problem at hand
you might be interested in the following reference :
http://www.marc-boulle.fr/publications/BoulleECML14.pdf
(random sampling approach in the lattice of the formulas generated by the flattening of relational data with a basic set of primitives)
.
Thanks for the answer, actually I already have the variables and arithmetic operators (e.g. +, *, /, etc) from a specific mathematical equation. My idea is I want to develop a new mathematical equation using the existing variables and arithmetic operators.
How big is the pool of arithmetic operators and variables? If it's human manageable pool size, one approach would be to build mathematic model by hand (using educated guess, leaner or not linear model etc) and then optimise parameter of a model by gradient descend methods given the loose function. If a pool size if unmanageable by hand, GP is the only way to go.
For equations development GP is the only choice here. Try RGP from R (https://cran.r-project.org/package=rgp) or HeuristicLab (https://dev.heuristiclab.com/trac.fcgi/). Both come with GUI and both are powerful tools. I personally prefer RGP. Good luck!
- Badges
- Science topic