Symbolic regression is a regression method that searches for both the form of a function and its numerical parameters.
Given data 
,
it explores a family of mathematical expressions 
and scores each expression using its prediction error,
often together with a penalty for expression complexity.
Ordinary parametric regression begins with a fixed form such as a polynomial of prescribed polynomial degree and estimates only its coefficients. Symbolic regression instead searches over trees of variables, constants, and allowed operations. A genetic algorithm is one common way to generate and modify candidates, although other discrete and continuous search methods can be used (Koza 1992).
Balancing fit against complexity is important because an unrestricted expression search can overfit the data. The result is often reported as a set of models describing different compromises between predictive accuracy and model simplicity.
