MIT artificial intelligence course Note

source: 网易公开课  patrick winston's MIT AI course

Lecture 1 

Big ideas

- MIT solve problems by Models

- Generate and Test is a powerful model (e.g. find a flower in a book by turning page one by one):  Generate and test method consists of generating some possible solutions, feeding them into a box that tests them, and then out the other side comes mostly failures.

- Rumpelstiltskin Principle says that once you can name something, you get power over it.

Lecture 2 (Goal Tree)  http://open.163.com/movie/2017/9/6/R/MCTMNN3UI_MCTMNUL6R.html 

---

Ask our self the nature WHEN YOU ENTER ANY NEW DOMAIN:

The nature of knowledge. What kind of knowledge is involved? 

knowledge about transformation, how goal tree works, what don't need to be transformed because you can lookup the table.

How's the knowledges represented?

a list of expressions

recorded in the table (integrals)

knowledge about goal trees embedded in a procedure (procedurally represented)

How is it used? 

transformations makes the problem simpler

integral tables is used to serve as the bottom of the tree

How much knowledge is required? 

table of integrals:  26

safe transformations: 12

heuristic transformations: 12

Interesting finding1: The method to be used and the characteristics of the problem is almost a diagonal table

They solve some problems, they seems smart, when they tell you how they did that, they don't seem so smart any more.