I want to achieve

In Optuna, what is the general guideline for the number of searches by changing parameters?
Of course, the more you do, the better parameters you will see, but as a guide, is there a number of times that you should search at least this many times?

There are about 5 types of parameters, such as the number of epochs, the number of intermediate layers, and the types of activation functions.
Since each parameter, especially the number of epochs, ranges from a small number to a large number, I think that the number of searches will be enormous even if there are few types of parameters.

  • Answer # 1

    There is no general search count as it depends on data, models and computational resources! That is the answer.

    That makes me lonely, so when I googled with "optuna" n_trials "", n_trials = 100 seems to be the majority. Since the tutorial of optuna is like that, I think that everyone is imitating. I couldn't find any description about how many should be.

    For the time being, try turning it with n_trials = 100, and it will be an about guideline to adjust up and down based on the learning time and the accuracy of the created model.

    By the way, LightGBM Tuner also has a tutorial of n_trials = 100, but since it is a search specialized for LightGBM, it will be exactly "things different" from n_trials = 100 of raw optuna.