Orbit correction is one of the most fundamental processes used for beam control in accelerators. Algorithms have been developed at various laboratories to meet specific demands. Typically, linear algebraic tools are applied to various response matrices to solve related problems. However, there are still many problems faced by orbit correction algorithms such as lengthy measurement and computation time. A new approach based on machine learning to develop an orbit correction program is introduced. In this method a machine learning program is trained with correctors data and BPMs data for applying to orbit correction. Mathematical formulation, algorithms prototyped and tested on simulated and real data, and future possibilities are discussed.