A novel method of using Machine Learning (ML) algorithms to improve the performance of Linear Genetic Programming (LGP) is studied. In this study, structures used to organize the trained ML models are called Experience Models (EM). They are used for different mutate actions of the mutation operator in LGP. The purpose of using EM is to regulate the random search performed by the mutation operator. The aim of using EMs is to let the suitable candidates have higher chances to be selected. In this study, two sources of knowledge are used to create the training sets that are used to train ML models. The first source is the pre-existing knowledge of symbolic regression. This knowledge reflects the effect of adding one math function segment to another math function segment. The second source is the knowledge generated during the evolution of LGP. This knowledge reflects the effect of using different gene components at different chromosome indexes on the overall fitness. Based on these two sources of knowledge, two types of EM are designed. They are Static Model (SM) and Dynamic Model (DM). The SM uses ML models trained with the first knowledge source. A SM tries to achieve the aim of using an EM by reducing the size of the candidate sets used by the increase action of the mutation operator. The DM uses ML models trained with the second knowledge source. A DM tries to achieve the aim of using an EM by creating distributions of gene component types, which can reflect the information in the second knowledge source, for change action of the mutation operator. In this study, SM is used only for increase action in the mutation operator; DM is used only for change action in the mutation operator. From the experiment results, if compared with a LGP, when a LGP using a SM, it tends to need fewer generations to have a hit, at the same time achieving similar mean best fitness. In contrary, when used with a DM, a LGP do not show performance improvements.