Breakout Bulletin: Pretesting the Effectiveness of Genetic Programming

[LGTrader]

Mike,
   I wanted to say thanks for the newest Breakout Bulletin. It's a set of interesting ideas put together in an engaging way. I've been playing with the ideas for the last few days with some interesting results and (apparent) improvement in my methods for generating models in Builder. Thanks for publishing the ideas.

   For any other group members who don't happen to be signed up for Mike's newsletter you can access the bulletin I'm referring to here:

http://www.adaptrade.com/Newsletter/NL-Pretesting.htm

   One potential new Builder feature comes to mind for me as I've been thinking about how to apply these ideas to actual model building vs the study portion (my wording) that the bulletin focuses on. Would it make sense to have a mode for the 'Reset on OOS Performance' feature that did the reset based on the average of all the models saved vs just the model with the highest fitness? It seems to me this would potentially aid in finding stronger _groups_ of ideas during a long Builder run instead of just focusing on the single model with the highest fitness and throwing the whole group away. This might address the frustration of users sometimes report of seeing what looks like an interesting model in the middle of a group of generations but losing that model because it dropped down and some less interesting model ended up with the highest fitness at the decision point causing the whole process to begin again.

   For clarity, in my mind this new option would be just a check box in the same Reset group feature, disabled by default with Builder operation today unchanged. If enabled then the decision to reset would be made on the average of the whole group.

   Again, thanks for the bulletin ideas. I hope it results in improvements for others as it appears to be improving my results.

Cheers,
Mark

[Keith]

I also would like reset on some choice other than the fittest strategy

Sent from my iPhone

[Michael R. Bryant]

Thanks for the suggestion, Mark.

 

Mike Bryant

 

Subject: Breakout Bulletin: Pretesting the Effectiveness of Genetic Programming