Today I begin the study of Reinforcement Learning, which is one of the interesting research field i would like to study. These few years we see the research team DeepMind has done a lot in this field, and made numbers of achievements. We see AlphaGo in GO, we see AlphaZero in multiple chess games, we also see DeepMind tried to create a human-like ai in the game Starcraft II. Although finally they didn’t successfully create an ai that behaves like human, maybe due to the limitation of current algorithms, ai architecture or other reasons. However, they also discover many interesting ai behaviors, like ai will lift the base on the sky, in order to avoid enemies’ attack. They are also some already made human-like ai in games like Super Mario using Reinforcement Learning.
For me, I am expecting one day i can create an immersive and powerful ai in some RPG games, like The Elder of Scrolls : Skyrim or Sim series. I am also expecting one day the NPC in games can be like a human or more crazy, we cannot even distinguish between the guy facing you is a real person or a NPC. That’s really cool!
Also, I am always thinking that, a good ai can be composed of some basic general principles, instead of combination of a vast number of special purpose tricks, procedures and heuristics. In these days, most of the ai is realized by finite state machine and behavior trees. I am not saying they are not good technologies, in truth they are very powerful in designing a good ai. But that’s not enough for an ai to behave like a human. We can try making a more complex ai, but that’s just adding more characteristics to an ai. We need some way to let ai have the ability to learn something, in the guidance of some general principles. That’s why i am interested in this field.
In the remaining of this article, I will write down what I learn and what I think after reading the book < Reinforcement Learning : An Introduction > by Richard S. Sutton and Andrew G. Barto.