Self-driving is one of the most potential valuable field of artificial intelligence. In order to realize safety driving on the real road, cars not only need to recognize and track the objects with its perceptual equipment, but also act properly according to the current road conditions with decision-control module. Among them, the decision-control module is the critical part, and also is the difficult part. It includes action prediction, obstacle avoidance, trajectory planning and so on. As a strong tool for dynamic modeling, finite state model is wildly used in the decision-making modeling of self-driving. This method summarizes the events which change the action of unmanned cars, and divide them into finite number. Each action represent the different ways to control cars, and build the decision-making model with rule-based or statistical method. Rule-based method can achieve functionality quickly, but it can’t list all the decision states and lack of ability to describe uncertainty information. Because of its weakness, it is essentially necessary to combine with statistical methods. In addition, with the advent of simulation engines like Carla and Torcs, reinforcement learning is gradually used in the decision-making research and make some achievement. However, reinforcement learning is mostly used in the visual game engine, and it is far from real road test. What’s more, the stability also need to be further strengthened.
We use bayesian network and bayesian nonparametric models to tackle the small data decision-making problem in self-driving research. Unlike traditional rule-based approach, these methods can take the casual correlation in time series data into serious consideration. They can not only fetch the features of the data effectively, but also significantly reduce dependence on the amount of data. Due to these qualities, they can improve the accuracy of small data decision-making comprehensively.