AI Course Study Hall

How can we expose more people to critical thinking?
User avatar
Rob Lister
Posts: 21510
Joined: Sun Jul 18, 2004 7:15 pm
Title: Incipient toppler
Location: Swimming in Lake Ed

AI Course Study Hall

Post by Rob Lister »

I"m just not getting this notation

Image

for the linear regression. I watching the cost function video and the whole thing was fucking greek to me. He was doing it as review. fcik.

is this function:

Image

simply defining the (learned) slope? or is there something more fundamental working?

User avatar
DrMatt
BANNED
Posts: 29811
Joined: Fri Jul 16, 2004 4:00 pm
Location: Location: Location!

Re: AI Course Study Hall

Post by DrMatt »

The opening formula defines a linear function h in one variable in terms of two fixed parameters.
The second formula appears to be defining J as an integral. If J is a cost formula for a computation, it's probably the integral of all past costs up to the present, or the integral of all subsidiary costs up to the current input-size, or some combination of both. If it's a cost formula for an NP-complete problem like the traveling salesman, it's probably summing up the costs of each step in a proposed solution.
Grayman wrote:If masturbation led to homosexuality you'd think by now I'd at least have better fashion sense.

User avatar
Mentat
Posts: 10271
Joined: Tue Nov 13, 2007 11:00 pm
Location: Hangar 18

Re: AI Course Study Hall

Post by Mentat »

Is this for the machine learning class, or ai class?
It's "pea-can", man.

Lapis Sells . . . But Who's Buying?

User avatar
Rob Lister
Posts: 21510
Joined: Sun Jul 18, 2004 7:15 pm
Title: Incipient toppler
Location: Swimming in Lake Ed

Re: AI Course Study Hall

Post by Rob Lister »

Mentat wrote:Is this for the machine learning class, or ai class?
Promise not to laugh?

User avatar
Mentat
Posts: 10271
Joined: Tue Nov 13, 2007 11:00 pm
Location: Hangar 18

Re: AI Course Study Hall

Post by Mentat »

Nope. :)
It's "pea-can", man.

Lapis Sells . . . But Who's Buying?

User avatar
Mentat
Posts: 10271
Joined: Tue Nov 13, 2007 11:00 pm
Location: Hangar 18

Re: AI Course Study Hall

Post by Mentat »

The thetas (Øs) are the parts an algorithm's guess for the solution. For now, assume that Ø is fixed. Then hØ is the application of the guess data into an actual function. That guess is just a model for the data (be it good or bad). The J function evaluates how accurate any guess can be. So you plug in your guess - the Øs - and you get a value. The point is to find a (preferably) global minimum, since J returns the sum of squares of errors (ie big value = lots of error, small value = little error).
It's "pea-can", man.

Lapis Sells . . . But Who's Buying?

User avatar
Rob Lister
Posts: 21510
Joined: Sun Jul 18, 2004 7:15 pm
Title: Incipient toppler
Location: Swimming in Lake Ed

Re: AI Course Study Hall

Post by Rob Lister »

Mentat wrote:...since J returns the sum of squares of errors (ie big value = lots of error, small value = little error).
That's what I was missing. tnx.

Mentat wrote:Nope.
Then I'm not going to tell you that I invested about 12 hours in learning this stuff because I clicked the wrong link and just dived in without noticing. (I got a good grade though!) Now I have to decide if I'm going to try to finish it.

don't tell ed. he'll make fun of me.

User avatar
Rob Lister
Posts: 21510
Joined: Sun Jul 18, 2004 7:15 pm
Title: Incipient toppler
Location: Swimming in Lake Ed

Re: AI Course Study Hall

Post by Rob Lister »

Homework Notes

http://larvecode.tumblr.com/tagged/ai-class

excellent summation of terminology. i was screwing the pooch on some of these, particularly Full and Partial Observable. I sucked ass on the homework -- counting the nodes --! for the different algorithms. I might go back and correct my homework.

Here's a tough question: would chess be full or partial observable.

clearly the theory is full observable but is there room to consider style? to considering how an opponent might more likely move given his history of moves?

User avatar
Mentat
Posts: 10271
Joined: Tue Nov 13, 2007 11:00 pm
Location: Hangar 18

Re: AI Course Study Hall

Post by Mentat »

Chess is fully observable. That there is an unpredictable opponent does not mean it's partially observable. It just means you can - given enough computation - map out and search every possible contingency at any time given any initial state. What would not make it fully observable is if there exists contingencies that you couldn't map out. Example: trying to drive through a city with no guidance (such as a map, GPS, or previous knowledge).

At least, that's how I understand it, I may be wrong.
It's "pea-can", man.

Lapis Sells . . . But Who's Buying?

User avatar
Mentat
Posts: 10271
Joined: Tue Nov 13, 2007 11:00 pm
Location: Hangar 18

Re: AI Course Study Hall

Post by Mentat »

FWIW, I did write a machine learning algorithm for the Sorry to Make You Cry game two years ago, and went from last to first after coming in 5 weeks late. Then quickly lost it as the program consistently gave guesses that were much worse than random selection in the last few weeks. Never fixed it as I was way too busy with school work at the time. :oops:

ETA: Here's a thread about it: viewtopic.php?f=5&t=29048&p=553082#p553082
It's "pea-can", man.

Lapis Sells . . . But Who's Buying?