The 2019 DeeCamp Artificial Intelligence Training Camp Achievement Show and Couture Ceremony was held at the Yanqi Lake Campus of the Chinese Academy of Sciences. In the past four weeks, the participants completed 50 AI-related projects initiated by 22 companies, and received a total of 115 from Counseling for technical leaders and engineers of the company. A team led by Tianhao Zhang from the laboratory successfully built a high-margin AI landlord product from 0 to 1, in just three weeks, and won the DeeCamp Best Team Award this year.

　　

LeiFeng.com and TianHao Zhang had a conversation, and this article intercepted some interviews.

Reporter: The original intention of participating in the training camp?

Tianhao Zhang: Last year, my roommate participated in DeeCamp. It was held at Peking University. At the water platform, I took a lot of professional courses. Especially the course of Kaifu teacher made me particularly impressed. Let me think about labor from the perspective of industry. The development of intelligence. Later, I paid close attention to the project of the roommate. It was about the robotic arm grabbing. I felt that the students could learn a lot from the project. So I paid attention to the registration this year and found out about the game AI project. The scientific research direction of my group robot intelligent control is relatively close, all about the intelligent body making decisions in the environment, so I want to learn.

Reporter: How do you team up and divide the work in the project?

Tianhao Zhang:We are randomly teamed up and we don't know each other before we team up. The captain is going to pick himself up later. Our group is very fortunate. Everyone has their own technical field, and they all come with a learning attitude. They all want to gain something in this summer camp. Through the study and discussion of the thesis, we found the research direction that each classmate is more suitable to break through in three weeks, and discussed and tried it efficiently from the beginning. I especially thank the team members for their trust in my project planning. We only have three weeks to make a project that everyone is quite satisfied from scratch, which is very tense in time. Because this project removes the algorithm aspect, there are still many engineering things to solve, such as the game engine, game server, database, game front end, game interface, music and so on. During that time, we probably averaged five hours of sleep per day. There may even be some rotations. After a classmate has finished an all-nighter, he goes to rest for five hours. Another classmate gets up and then proceeds to make some adjustments, and then the two will do some more discussion. Except for the open days provided by the Innovation Workshop, everyone basically did not go out to play.

　　

Reporter: Can you explain to us your model practice in detail?

Tianhao Zhang:We innovatively proposed a multi-model fusion Landlord AI framework, first designed a number of landlord AI models, and through offline learning to make them each have a certain landlord ability, and finally through reinforcement learning to do Multi-model fusion, and finally choose the decision result of one of the models.

The algorithm framework is as follows:

　　

Reporter: How effective is your model?

Tianhao Zhang:Each of our models has achieved a certain landlord ability, and our enhanced learning model is based on the same baseline confrontation, achieving a higher winning rate than the existing paper. The second is to supervise and learn the model that imitates the behavior of human players. It also achieves a prediction accuracy of 76.5% on the test set. This is only a model trained by 35W data in gold segmentation. If more data is obtained, Whether the effect will be better, we don't know. However, it is obvious that this AI does have some imitation behavior.

Reporter: What is your biggest harvest and feeling this time?

Tianhao Zhang:My harvest exceeded expectations. At the beginning, my expectations were algorithms and how to solve decision problems. However, DeeCamp is not just about algorithms, it lets us know how everyone should work together to complete a project and complete a real product. I am very fortunate that there are all kinds of talents in our team. We need to discuss and communicate every day. Coincidentally, our team members are all in one bedroom. Therefore, every two nights we will have a small meeting, summing up in five days, watching the progress, looking for a combination, and then arranging the players to do some crossover and technical docking. This may be more like doing something in the team. After the project starts, we need to understand what other people are doing. I think these things are very important for future business work. You need to know how to work with people, what they want, what you can offer, what you want, what you need them to provide, and this is also very important.

　　

Reporter: How do you balance your studies and the time of the competition? Does the supervisor support it?

Tianhao Zhang:This is a very difficult thing, because the doctoral student's scientific research task is relatively heavy, and the teacher will worry that my participation in DeeCamp is not as good as research in the laboratory. This month has delayed the progress of one project and two papers in the lab. In addition, since the undergraduate students of Peking University will study in the laboratory during the summer, I can only discuss the project with the undergraduates who are doing research with me remotely. . Our current research is a cross-cutting direction. We will have some math, physics, automation, mechanics and computers. The teacher hopes that I will not go to the pure computer direction. He thinks that this will lose my advantage. I also recognize the teacher. the opinion of. Therefore, I am especially grateful to my tutor Xie Guangming for his support and thanks to the lab to help me share the work of the original laboratory.

Reporter: What is your research plan next?

Tianhao Zhang:I am doing multi-underwater robot group control, which involves the perception and control of water surface and underwater, combined with AI surface object recognition, underwater image enhancement, robot sensor data processing, robot robustness control, etc. Nowadays, land resources have been developed almost, and the ocean has not yet developed. In addition, the country has also put forward the slogan of "marine power", so the research on underwater robots is related to life and the country. Our laboratory has developed a number of underwater robots, which have also assisted national expedition personnel to do some exploration in the Arctic and the Arctic. They can also have many applications in fisheries, water quality monitoring, and rescue.

This chapter is selected from LeiFeng.com. The original title is “Sacrificing Sleep, Delaying Papers, Developing the Landlord AI in three weeks, they won the DeeCamp Best Team Award this year”, author skura, if you need to reprint, please apply to LeiFeng.com Authorization.

Address   https://www.leiphone.com/news/201908/fxbKOrgvOni8WVAy.html