Avoiding Overfitting in Symbolic Regression Using the First Order Derivative of GP Trees

Genetic programming (GP) is widely used for constructing models with applications in control, classification, regression, etc.; however, it has some shortcomings, such as generalization. This paper proposes to enhance the GP generalization by controlling the first order derivative of GP trees in the evolution process. To achieve this goal, a multi-objective GP is implemented. Then, the first order derivative of GP trees is considered as one of its objectives. The proposed method is evaluated on several benchmark problems to provide an experimental validation. The experiments demonstrate the usefulness of the proposed method with the capability of achieving compact solutions with reasonable accuracy on training data and better accuracy on test data.