Regularization adds to the loss function a penalty on the size of the coefficients, in order to limit the model's complexity and overfitting. The strength is set by a hyperparameter (alpha for Ridge and Lasso). Variables generally need to be standardized beforehand.