Evolutionary Algorithm, that can optimize the choice of hyperparameters and architecture of machine learning models.

• in folder "experiments" are all executable files
• use runner.py from the root directory to run the experiments
• define a config.py with the global vars CHAT_ID and BOT_KEY, if you want to use send_telegram

Notes:

• a small unnoticed difference in the instance is enough for the difference, the cross instances, of the very similar instances are again very same!
• Solution:

  • only max one cross instance

  • for crossing minimum dissimilarity, not just stupid the best ones

  • Metric for similarity

    • n_duplicates in generation
    • n_duplicates only one difference

  • crossing of two individuals, that are only different in the same position, the duplicate protection loop will go forever!

  • even if they are not the same, they must differ in 2 attributes

Ideas:

• if there is no better individual for many generations, the generation could get more random -> adaptive search: instruments provided by DNA! metric, when the new generation was good enough, in crossing, similar values will appear most of the time
• diversity good over attributes?
• How to evolution.
  • you need a very precise fitness function!!! KFold required!
  • the derivation between Folds should be low
  • most importantly the differnce between KFold mean and the final training should be low -> representative 70 (derivation 1 percent) mean to 72 69 (over 10 percent derivation) mean to 65
  • write a function the does the check and has as output a good metric that you can see, if an optimization of your fitness function will lead to a better model!!!!
  • cross could be an individual in between
  • no better individual for some generations, the change rate needs to be both bigger and smaller - smaller for micro optimization, bigger for finding a complete new optimum
  • finding another optimum, I could open a new familiy tree, where the maximum is not compare with my main generation