There have been many attempts to solve the decision-making problem for the driving behaviors of autonomous vehicles. Early attempts employed empirical human knowledge to define rules explicitly, and thus limited by the diversity of human experience and the depth of human thinking. Recent attempts built neural network models based on Computer Vision and Imitation Learning, and thus vulnerable to unexpected behaviors caused by the black-box system of the model. In this paper, we present an alternative approach to solving the decision-making problem while avoiding the issues above using Program Synthesis. Our model generates human-readable programs to represent the necessary and sufficient conditions for safely executing some chosen driving behavior, and thus significantly increase the model transparency to avoid accidents. At the same time, it is also powered by the virtually unlimited machine learning capacity. Along with our program generation pipeline that generates programs from driving scenes of some chosen driving behavior, we also design a driving simulator that’s capable of generating diverse driving scenes efficiently, and a domain-specific language to describe all the conditions we account for. So far, we have built a solid foundation for our approach, and we will continue to tune our model and validate our pipeline on more scenarios.