In this dissertation, we attempt to improve operation at particle accelerator facilities by applying various machine learning techniques to several key areas of acceleration operation for a particle accelerator, such as beam orbit control, beam quality measurement, and beam optimization. For complicated large scale machines such as accelerators, the inner mappings between control parameters and performance data are often only partially known, but they can be learned and simulated using machine learning methods without extensive physics understandings. In this dissertation, two major types of machine learning techniques were developed for and demonstrated on several accelerators located at Cornell University and Brookhaven National Laboratory: neural network design for constructing accurate mappings between control parameters and performance data, and optimization algorithm development for finding optimal operation parameters automatically. These successful applications show the benefits of integrating machine learning algorithms with accelerator control system, and build the foundation for including similar techniques in the ongoing development and construction of the Electron Ion Collider at Brookhaven National Laboratory.