help -best option about parameter learning!

hello I have a question regarding the learning of the parameters, which would be the best method to use in an implementation never, always, selective, I think that by discarding never is not the best option, my doubt is whether to use always or selective, if I use I will always learn all the parameters that you send in the requests, it is recommended that you learn all the parameters by strain, I understand that if I know how well the policy could eliminate the willcard? , or use select since you only learn the parameters that exceed the willcard, how to define the properties of the willcard so as not to suffer an overflow attack?