Open mrektor opened 5 years ago
Hmm, thank you for posting this idea here. I should be easy to implement since TPOT can record those statistics during optimization. We will look into the paper for more details.
Sounds like an interesting research project and relates to the 'age-layering' concepts that we briefly explored in the early days of TPOT. The feature would have to be thoroughly evaluated prior to implementation into the dev/master branch of TPOT, given its far-reaching consequence on the software.
Since TPOT is a GP framework, an interesting feature would be have the possibility of reguralize its evolution with a lifetime of the pipeline. That would mean that a pipeline can only survive "n" generation and then 'die'.
Context of the issue
This technique was proposed in a recent google paper where they used evolutionary programming for automatically designing a deep neural network achitecture (that eventually became the actual state-of-the-art in various computer vision problems).
This technique end up in keeping not "good individuals" that have had a perhaps noisy good value on a particular dataset, but rather a good ancestral code that can last the test of time. This was a somewhat poetic description, but the paper is much more detailed about: https://arxiv.org/pdf/1802.01548.pdf
It should not be too difficult to implement.
What do you think about it?