test and train correlation coeff

[mehdikazemi1995]

hi dr bhrnjica i am using GPDOTNET version 5.1 in particular project. the training correlation coefficient is 0.987 but the testing correlation coefficient is about -0/03!!! how should i change the genetic programming parameters to increase the testing correlation coefficient?

[bhrnjica]

Hi @mehdikazemi1995 as you probably know your model is overfitted.

You have several options in order to get better prediction:

• check if you normalized data by one of two available methods in GPdotNET
• reviewed you features and see if you can reduce them
• Review Function Set, try to decrease them. Try with only few of them (*,/,-,+, thanh)
• decrease the iteration number
• increase population number
• decrease tree depth during crossover and /or mutation

Hope some of them can help you to get better model.

[mehdikazemi1995]

Thank you so much, I will check the ways out