The Facebook researcher Yuandong Tian is also a famous answerer in Zhihu, a Chinese most popular Q&A platform as Quora to the U.S. So I am very excited to see him in person.
This topic is very interesting because games and artificial intelligence are all my favorite fields. Therefore combing them is quite attractive to me.
In the lecture, Tian started from why game is so charming and explained what they’ve done in training game AI agents and ended up with how those agents performs comparing with traditional rule based agents.
The structure of the content is quite skilled and logical. The beginning is very easy for audiences to get involved and the difficulty enhanced increasingly. Although at the end of the lecture, even the codes were shown, which made the lecture hard and boring, when the examples of demo are played the lecture is still attractive and interesting. This inspired me that what attract the audiences may not be the content, but could be the method of presentation. We can attach some interesting part to the hard part.
AI shocked the world when AlphaGo defeated Lee Sedol in the recent 2 years and it is quite hopeful that AI can beat human beings in other games. For example, the agent from OpenAI defeated world campion in Dota this year, which give us a hint that it is becoming stronger and more intelligent.
Now most of the game AI agents are just rule based which means they lack changes and adjustments. Even though they have a lot of strategies, their reaction are predictable i.e., once we find the rule, we will always win.
However, after training by deep learning, AI will be more like expert human players and players will find the game more challengeable.
Although the agent react well in the video, I am more interested in the challenges it is facing.
1. About data
We all know to learn about the world, AI needs more information than humans to understand concepts or recognize features. In the lecture, the agents are trained by capturing the screen images, we need to let AI understand everything and recognize every feature in the image from any scopes, any angles. We need to find better ways to make a evolution to make our deep learning systems more efficient and able to work with less data.
2.About change
In gaming, the scenery may be changed very sharply, if training the agents takes too much time, it is not quite efficient, therefore, a flexible model is needed. In the neuro network, maybe we can have less layer or nodes in order to prevent overfitting and give AI the ability to accommodate to the change of environment.
3.About training
Now the training is on the supercomputer, I think in the future, when the complexity of training is lower and the calculating power of our PCs or cell phones are higher, we can make our private devices to train the model based on how individual performed in the previous rounds of the game. This is because different users have different preference, to adjust to different user, the AI agent will satisfy more users requirements.
To sum up, this lecture is organized very well, audiences can preview the this interesting combination in a nutshell. I am also inspired by the structure of the presentation: different level of audience can get what their level can understand in the lecture and never fell boring. As my field of study is also about AI, I will keeping focusing on the following news from this team and read their papers.
