In recent years, the artificial intelligence technology represented by deep learning and reinforcement learning has expanded from the engineering technology fields such as language translation, image recognition and industrial automation to the economic and financial fields such as intelligent production, intelligent agriculture, intelligent logistics, big data macroeconomic monitoring and quantitative investment research. It can be said that it is widely used. Artificial intelligence technology has the inherent advantage of processing high-dimensional data. It can avoid many limitations of traditional analysis methods by means of representation learning, value function approximation, feature selection, etc., and obtain better prediction and decision-making results. In order to make artificial intelligence technology achieve satisfactory prediction and decision-making results, researchers often need to invest a lot of data resources. This technical characteristic makes data resource become a key production factor. Under the background of the increasing popularity of big data, intelligence, mobile Internet and cloud computing, artificial intelligence technology, as the underlying technology for providing information products and services, is also the key to the gradual transformation of industrial economy to digital economy. What are artificial intelligence algorithms? Artificial intelligence algorithms can be divided into supervised learning, unsupervised learning and reinforcement learning. Among them, supervised learning is to learn the rules from human experience through continuous training programs (models). In this kind of machine learning, researchers will constantly adjust the model parameters to achieve the learning purpose by marking the data. It is similar to that parents will show children apples of different colors, sizes and even types, and teach children to know "never seen" apples. This is the purpose of supervised learning: out of sample prediction. Unsupervised learning makes the machine directly extract features from the existing data through the training program, compress the information and use it to complete other tasks. As in the traditional principal component analysis, high-dimensional features can be approximated by low-dimensional vectors. For example, we can use principal component analysis technology to compress images to save storage space. Therefore, this kind of machine learning algorithm does not need previous experience and is also called unsupervised learning. Of course, unsupervised learning and supervised learning are not opposed to each other. For data with partial annotations, we can also use semi supervised learning algorithm. For example, the recently popular adversarial neural network - we can use this algorithm to learn a series of Oracle Bone Inscriptions and make it generate a number of "oracle bone inscriptions" that are enough to confuse the real with the real, but never represent any meaning, which is equivalent to the calculation program "drawing a tiger from a tiger" without knowing that it is a tiger. In addition, reinforcement learning is different from the above (none, half) supervised learning algorithm. Reinforcement learning is an extension of dynamic optimization, while (none, half) supervised learning is closer to statistics. Reinforcement learning maximizes the cumulative benefits of intelligent programs by making them interact with the environment and adjusting the decision-making parameters (processes) of intelligent programs. Reinforcement learning is a machine learning algorithm closest to human decision-making process. It is similar to enabling an agent to perceive the world infinitely and quickly, and optimize its decision-making process through its own failure or success experience. In this process, computer programs do not need teachers so much. Of course, reinforcement learning cannot be completely separated from supervised learning. For example, alphago is a computing program trained by reinforcement learning,