Need to provide ERCs

Created by: Vaguery

Success in most symbolic regression and classification problems depends on the use of Ephemeral Random Constants as part of the solution library. As far as I understand it, in this version one has to explicitly provide specific constants, as part of the training data set.

As things stand, even with the availability of user-provided specific constants, any search that requires numerical optimization of coefficients is spending most of its effort on constructing numerical arguments, rather than actually fitting qualitative curves or decision planes.

A simple example:

One of the pedagogic examples I always use is the birthday polynomial. Given your own birthday in year month-number day-number format, and then attempt symbolic regression on a dataset of

y = year + month-number * x + day-number * x * x

Without ERCs, symbolic regression on this simple polynomial spends most of its effort trying to construct large numbers from units. The interplay between this "sub-task" of actually constructing large constants from small ones, and the more important task of explaining the data (which is usually the intent of the user) can lead to trouble, since from the standpoint of the GP system itself there is no differentiation between these tasks.