Hi, I know this isn't strictly a game programming questions, and for that I'm sorry. I'm in a bit of a crunch to get a project done for school, and I'm looking for a machine learning algorithm.. but there's alot out there, and I don't have the time to read many papers or books. I figured this would be a good forum to ask about something like this in.. plus I don't know of any others. :( And as you can see I'm a new poster, though frequent reader. Here's the thing I need.. if anyone has used Quicken, you'll know what I'm talking about: When you set up something on a regular basis, say to pay a bill every month. After a few months, Quicken will let you know it has noticed that you do this regularly, and then offer to start doing it automatically for you. Thats the kind of thing I want.. My idea is to have something keep track of usage pattern on a specific app on a PDA, then after it can predict the usage with a certain degree of confidence, to ask you if it can start doing it for you whenever you start that action. I don't need to acutally implement this, but the name of the algorithm so I can describe it in my report... plus I need to read up on it too so I know what I'm talking about. :P Some things I've started looking into are reinforcement learning, and also the Q-learning algorithm. Is that sort of the right thing to be reading, or is there another algorithm that's really tailored for this sort of thing. Thanks alot! David

Have also a look into Evaluating Hypotheses and maybe also
GA might be a way to solve Reinforcement learning...
The problem here is seems a sparse data problem to me though ...

Hi snufer,

Thanks.. What is GA?

I'll start looking for evaluating hypthesis and spare data information as well.

David.

Quote: Original post by mwa
Hi snufer,

Thanks.. What is GA?

I'll start looking for evaluating hypthesis and spare data information as well.

David.

GA is kown as Genetic Algorithms is quite useful many things. Even for
Game programming.

Check out this site for a very
good introduction into GA and its realization...

P.S: I am kinda in hurry right now. Give let's say 3 hours and I explain you
a possible solution to your problem. (Just a possible soution) Alright?

See ya!

snufer,

sounds good!

thanks for the link.. though a GA seems like a bit of overkill, no?

and any possible solutions or ideas would be greatly appreciated.

David.

Quote: Original post by mwa
snufer,

sounds good!

thanks for the link.. though a GA seems like a bit of overkill, no?

and any possible solutions or ideas would be greatly appreciated.

David.

Does it need to be some topic out of Machine Learning?
Or do you just need to sell it as it has been done with Machine Learning
concepts [lol]

Basically you can just evaluate hypotheses. i.e: During the first few
days just record any new closing data. (This could be a date or something)
Stack all the collected values. And try to approximate a function...

just like that. (Well hopefully better then that) [grin]
Well how do you achieve that?
Assume every time you request an action which might be an action with
some frequency in it, you basically confirm the action as well just by requesting it. So this is a very important messure for you to evaluate the hypothesis.
One approach you might want to consider is someting like that:

PLAN A)
All raised measurements must be, within a certain range of tolerance around
the expected value. For the time being we'll limit ourselves to
a Test with the help of the arithmetic mean to an average sample.
Fast the question pops up how the measured arithmetic means are distributed
around the theoretical expected value. (i.e: Gaussian, Normal, Binominal
etc..). Hereby it is important if the empirical variance of the basic population is know or if it needs to be guessed by the average sample.

Example for known variance:
When you roll the dice let's say n times you can calculate an average number.
This should be araound 3.5. But it surely can deviate from the average.
i.e: 3 or even is possible. (Depending on n). The average number though is on its part a variate. And the question now is, how is its distribution...
For any forumula you should consult a good book or ask your tutors about some
statistics. Especially if the variance is not known!!! And I think this is
exactely your case.

PLAN B) ;-)
If you don't want to put up with statistics although it this case I think
it's the best suited solution, is to apply some exotic stuff. I.e:
-GA and Neural Networks

GA:
Well I thought about GA here but I must confess It is very hard to code
that problem. Well I finally came up with an idea. The core problem I had was
to determine a good fitness score!
My solution: Record a couple (i.e: 10) measurements. And then calculate the
average value for all values. (as simplicty) Take the two best suited values
and crossover and perform some mutation over them... And so on...

Neural Network:
An other solution is to use A Kohonen Neural Network. Not a regular neural network though. More information click linky

I hope this post was any help to you. If you have any issues just PM me

[Edited by - snufer on April 12, 2005 5:15:04 PM]

hi snufer,

haha.. no i don't suppose it has to be strictly machine learning.. however i figured it would be nicer if i could use a somewhat known machine learning algorithm or concept, something i could name and maybe get a paper to reference as well, rather than just put my own solution with assurance that it's the best thing to do.

David

Quote: Original post by mwa
hi snufer,

haha.. no i don't suppose it has to be strictly machine learning.. however i figured it would be nicer if i could use a somewhat known machine learning algorithm or concept, something i could name and maybe get a paper to reference as well, rather than just put my own solution with assurance that it's the best thing to do.

David

you know I knwo the tutors sometimes [lol]
Sometimes they're nuts and bullheaded and read something about machine
larning and say OH YEAH MACHINE LEARNING IT IS! [grin]

Bayesian classifiers might fit your need. Just kind of like how the windows start menu and many microsoft products do folding in the menus to hide things you don't use often. Except here, you simply make a choice as to what macros to show based on some probability of usage. Simple and straight forward.

Thanks Weirdo (lol),

I'll look into that as well.. and yeah, the way windows tries to be helpful is a good example of what I'm trying to do. :)

David.