- On Sun, Jul 8, 2012 at 12:38 AM, Amit Aronovitch <aronovitch@...> wrote:On Sun, Jul 8, 2012 at 12:28 AM, Amit Aronovitch <aronovitch@...> wrote:On Sat, Jul 7, 2012 at 9:06 AM, Om er Zak <w1@...> wrote:
On Sat, 2012-07-07 at 08:20 +0300, Ori Idan wrote:

[... description of the problem was snipped ...]> My wish is to have an algorithm which requires only bounded

> number of

> computations of random numbers (i.e. no loop) to accomplish

> the above.

> Assume that N is large, so keeping lists of questions in

> memory is ruled

> out.

>

(rather long, but nicely written and not hard to follow)

Whoops, missed the "N is large" thing (so O(n) memory might be too high, even if you just keep weights...). Still, a nice read to share.

I guess there's no escape from trading some memory for the time improvement...

sent before completed (whoops again)...

I ment: you do not have to store weights of ALL the questions, just for the questions the user had seen (and you have to store a list of these anyway). Since you know how many the user had already chosen, and the sum of their weights - first choose (using a single dart) whether you want a question from the seen set, or the unseen one (which is supposedly much larger). If the larger is chosen, choose uniformly. If the smaller (stored) set is selected, choose using the alias method or some simpler alternative.

- On Sat, Jul 07, 2012, Omer Zak wrote about "Re: [hackers-il] Algorithm for randomly choosing a question":
> The proposed approach has the problem of memory consumption of order of

Omer, here is an O(1) *RAM* algorithm. Note that you still have O(N)

> O(N) (where N is the number of the questions). I'd like to see an

> algorithm whose memory consumption is closer to O(1) than to O(N).

disk usage, of course - there are N questions, and they need to be saved

somewhere, and so is the history of the success for them.

If you drop the O(1) memory requirement (which I think is an overkill for

any type of modern computer), you can make things simpler.

Note - I assume the number of probability classes (in your example,

1, 1/2 and 1/100, if I remember correctly) is constant (O(1)) and not

O(N).

Keep on disk a file for each probability class - one list of questions

with normal probability (1), another list of questions with half

probability, and a third list of questions with 1/100 probability.

Each will be a list of numbers (indexing questions stored in a separate

file).

In *memory*, keep the number of questions in each class: n1, n2, n3,

and the partial probability sums:

p1 = n1*1

p2 = n1*1 + n2*(1/2.)

p3 = n1*1 + n2*(1/2.) + n3*(1/100.)

Each time you want to draw a random question, pick a random number

between 0 and p3. If it's between 0 and p1, you decided to take a

question from class 1. If it's between p1 and p2, you decided to take

a question from class 3. If it's between p2 and p3, then from class 3.

You decide *which* question to take in the obvious way - e.g., if you

got a random number p between p1 and p2, then the question number in the

class 2 list is m=(p-p1)/(1/2). You then need to look at the m'th

position in the on-disk list of class 2 questions, to get the actual

index of the question.

Finally, when the user answers you need to update your data. You need

to remove the question from the previous class (to do this in O(1) time,

just move the last element of the list into the newly formed hole),

and to add it to the new class (put it last), and of course update

n1,n2,n3 and p1,p2,p3.

Again, if you don't mind O(N) memory usage - actually just N*4 bytes -

then I suggest that you do keep the list of questions - just their

numbers - in memory. Then you won't need to read and write the disk

all the time.

Nadav.

P.S. This is of course a variant on the classic shuffling (or random

permutation) algorithm. In the shuffling algorithm, you need to draw

random questions from the list, but after you used a question, you don't

want to see it again (its probability becomes zero). Instead of keeping

the used questions on the list (and getting O(N^2) time complexity for

the whole shuffle operation), the trick is to move them to the end of

the list, and to remember how many questions we still have (in the

beginning of the list) to draw from.

--

Nadav Har'El | Sunday, Jul 8 2012, 18 Tammuz 5772

nyh@... |-----------------------------------------

Phone +972-523-790466, ICQ 13349191 |If Barbie is so popular, why do you have

http://nadav.harel.org.il |to buy her friends? - Hello Nadav,

Thanks for your answer.

I ended up implementing something similar to your proposal.

The questions are stored in a SQLite database, so I don't mind at all

O(N) disk space consumption. I did want O(1) RAM consumption, as RAM is

to be considered as a scarce resource in smartphones.

The number of probability classes is finite (3 if to be exact). In

principle, it is possible to have a more sophisticated algorithm for

selecting the next question (with unlimited, time varying probability

classes) but I don't think that the subject matter warrants this kind of

sophistication.

The disk files for each probability class are being kept implicitly - as

result sets of queries for each probability class.

The actual performance of the algorithm (O(1), O(N) or whatever)

actually depends upon the DB performance, which I didn't bother to

benchmark or fine tune as a function of the number of questions.

I did not (even implicitly) implement the optimization that you

suggested for O(1) disk filee modifications.

--- Omer

On Sun, 2012-07-08 at 10:23 +0300, Nadav Har'El wrote:

> On Sat, Jul 07, 2012, Omer Zak wrote about "Re: [hackers-il] Algorithm for randomly choosing a question":

> > The proposed approach has the problem of memory consumption of order of

> > O(N) (where N is the number of the questions). I'd like to see an

> > algorithm whose memory consumption is closer to O(1) than to O(N).

>

> Omer, here is an O(1) *RAM* algorithm. Note that you still have O(N)

> disk usage, of course - there are N questions, and they need to be saved

> somewhere, and so is the history of the success for them.

> If you drop the O(1) memory requirement (which I think is an overkill for

> any type of modern computer), you can make things simpler.

>

> Note - I assume the number of probability classes (in your example,

> 1, 1/2 and 1/100, if I remember correctly) is constant (O(1)) and not

> O(N).

>

> Keep on disk a file for each probability class - one list of questions

> with normal probability (1), another list of questions with half

> probability, and a third list of questions with 1/100 probability.

> Each will be a list of numbers (indexing questions stored in a separate

> file).

>

> In *memory*, keep the number of questions in each class: n1, n2, n3,

> and the partial probability sums:

> p1 = n1*1

> p2 = n1*1 + n2*(1/2.)

> p3 = n1*1 + n2*(1/2.) + n3*(1/100.)

>

> Each time you want to draw a random question, pick a random number

> between 0 and p3. If it's between 0 and p1, you decided to take a

> question from class 1. If it's between p1 and p2, you decided to take

> a question from class 3. If it's between p2 and p3, then from class 3.

>

> You decide *which* question to take in the obvious way - e.g., if you

> got a random number p between p1 and p2, then the question number in the

> class 2 list is m=(p-p1)/(1/2). You then need to look at the m'th

> position in the on-disk list of class 2 questions, to get the actual

> index of the question.

>

> Finally, when the user answers you need to update your data. You need

> to remove the question from the previous class (to do this in O(1) time,

> just move the last element of the list into the newly formed hole),

> and to add it to the new class (put it last), and of course update

> n1,n2,n3 and p1,p2,p3.

>

> Again, if you don't mind O(N) memory usage - actually just N*4 bytes -

> then I suggest that you do keep the list of questions - just their

> numbers - in memory. Then you won't need to read and write the disk

> all the time.

>

> Nadav.

>

> P.S. This is of course a variant on the classic shuffling (or random

> permutation) algorithm. In the shuffling algorithm, you need to draw

> random questions from the list, but after you used a question, you don't

> want to see it again (its probability becomes zero). Instead of keeping

> the used questions on the list (and getting O(N^2) time complexity for

> the whole shuffle operation), the trick is to move them to the end of

> the list, and to remember how many questions we still have (in the

> beginning of the list) to draw from.

--

Sent from a PC running a top secret test version of Windows 97.

My own blog is at http://www.zak.co.il/tddpirate/

My opinions, as expressed in this E-mail message, are mine alone.

They do not represent the official policy of any organization with which

I may be affiliated in any way.

WARNING TO SPAMMERS: at http://www.zak.co.il/spamwarning.html - Hi!On Sat, Jul 7, 2012 at 7:53 AM, Omer Zak <w1@...> wrote:
I want the application to be able to select any question randomly, but

with different weights, so that:

- A question the user never saw has weight of (say) 0.5.

- A question the user answered wrongly has weight of 1 (i.e. such a

question is twice more likely to be presented than a question never

seen).

- A question the user answered correctly has weight of 0.005 (i.e. it is

100 times less likely to be seen again than a question never seen

before).Keep three lists with pointers to "unasked" "wrong" and "correct" questions, moving questions between them.First choose from which list to pick your question (abuse length of list * weight for each list.)Than pick up a random question from your chosen list.if RAM is a problem, keep the lists offline :-)Udi.