LFD Book Forum (http://book.caltech.edu/bookforum/index.php)
-   Chapter 1 - The Learning Problem (http://book.caltech.edu/bookforum/forumdisplay.php?f=108)

 canon1230 04-08-2012 01:32 AM

If an urn contains 100 green or red marbles, and you sample 10 and they are all green, what is the probability that they are all green?

 htlin 04-08-2012 03:30 AM

Quote:
 Originally Posted by canon1230 (Post 1027) If an urn contains 100 green or red marbles, and you sample 10 and they are all green, what is the probability that they are all green?
The short reply is that probability does not tell us this answer, unless there are more assumptions. :clueless:

Put it to the simplest case, if you flip a coin and get a head, what is the head probability of the coin? The answer can be "any non-zero number" but there is no further information to pin it down.

Hope this helps.

 GraceLAX 04-08-2012 04:26 AM

Quote:
 Originally Posted by canon1230 (Post 1027) If an urn contains 100 green or red marbles, and you sample 10 and they are all green, what is the probability that they are all green?
A physics professor friend wrote this up to help clarify a common statistical misconception.
Do these help? Read through the second one for a similar example.

 canon1230 04-08-2012 09:44 AM

P[|Ein - Eout| > epsilon] <= 2e^(-2 * epslion^2 * N)

Since Ein = 0, N = 10, setting epsilon to 0.5, the inequality gives us:

P[Eout > 0.5] <= 2e^(-5) = 0.013+

This seems to be saying something nontrivial about Eout.

 htlin 04-08-2012 04:01 PM

Quote:
 Originally Posted by canon1230 (Post 1042) Does the Hoeffding Inequality allow us to say something about this probability? P[|Ein - Eout| > epsilon] <= 2e^(-2 * epslion^2 * N) Since Ein = 0, N = 10, setting epsilon to 0.5, the inequality gives us: P[Eout > 0.5] <= 2e^(-5) = 0.013+ This seems to be saying something nontrivial about Eout.
The in Hoeffding is subject to the process of generating the sample (i.e. ), not the probability on . Indeed it tells us something nontrivial (and that's how we use it in the learning context), but it does not answer your original question.

The question that got answered by Hoeffding is roughly

"What is the probability of a big-Eout urn (many red) for generating such an Ein (all green)?"

not

"What is the probability of Eout being small in the first place?"

The answer to the latter question remains unknown, but even so, we know that having a big Eout is unlikely because of Hoeffding.

Hope this helps.

 rukacity 04-10-2012 08:19 AM

I would think this way:

if p is probability of the outcome then for 10 trials there is p^10 probability of getting all favorable outcome.

 jsarrett 04-14-2012 12:00 AM

Hoeffding's inequality has a free parameter in your question, namely . It lest you say that since you have 10 samples, if you want to be 80% sure of the distribution of marbles in the jar (), then the jar is at most percent different from the sample.

where

simplifying:

so we think that with 80% confidence, the jar is at least 10.2% green.

We probably need a bigger N!

 pventurelli@gotoibr.com 04-14-2012 07:24 PM

In the second lecture, the Professor asked a question about flipping a coin:
What is the probability of getting all ten heads if you flip a coin 10 times count the number of heads, and then you repeat the experiment 1000 times.

The answer he gave was 63% -- I would like to know how this was computed.

Any help would be greatly appreciated.

Thanks!

 yaser 04-14-2012 10:11 PM

Quote:
 Originally Posted by pventurelli@gotoibr.com (Post 1287) In the second lecture, the Professor asked a question about flipping a coin: What is the probability of getting all ten heads if you flip a coin 10 times count the number of heads, and then you repeat the experiment 1000 times. The answer he gave was 63% -- I would like to know how this was computed. Any help would be greatly appreciated. Thanks!
The probability of getting 10 heads for one coin is (10 times) which is aprroximately .

Therefore, the probability of not getting 10 heads for one coin is approximately .

This means that the probability of not getting 10 heads for any of 1000 coins is this number multiplied by itself 1000 times, once for every coin. This probability is therefore .

This is approximately since . Numerically, .

Therefore, the probability of this not happening, namely that at least one coin of the 1000 coins will give 10 heads, is 1 minus that. This gives us the answer of approximately 0.63 or 63% that I mentioned in the lecture.

 pventurelli@gotoibr.com 04-15-2012 05:51 PM

Thank you Professor, that was extremely helpful.

 ivanku 04-18-2012 12:59 AM

Quote:
 Originally Posted by yaser (Post 1288) This is approximately since . Numerically, .
Is it correct to assume that from the coins tossing simulation in homework 2 assignment one should be close to that limit (where we're tossing 1000 coins 10 times and is the fraction of heads obtained for the coin which had the minimum frequency of heads)?

 elkka 04-18-2012 03:49 AM

The law of big numbers states that the average $\nu_min$ is close to the $E{\nu_min}$.

$E\nu_min$ can be calculated directly for this experiment.
$P(\nu_min=0)$=0.623576
$P(\nu_min = 0.1)$ = 0.3764034
$P(\nu_min = 0.2)$ = 0.00002;
and $P(\nu_min>=0.3)=0$ for the purposes of calculating the mean.

Therefore, $E(\nu_min)$=0.037644, and the average proportion of heads for c_min should be close to this number.

 SamK52 04-18-2012 06:43 AM

Allow me to format your post:

Quote:
 The law of big numbers states that the average is close to the . can be calculated directly for this experiment. and for the purposes of calculating the mean. Therefore, , and the average proportion of heads for should be close to this number.

 elkka 04-18-2012 11:26 AM

Quote:
 Originally Posted by SamK52 (Post 1399) Allow me to format your post:

 yaser 04-18-2012 01:57 PM

Quote:
 Originally Posted by elkka (Post 1413) Please, allow me to ask how you did it?
http://book.caltech.edu/bookforum/sh...77&postcount=1

 rohanag 04-18-2012 01:58 PM

how did you calculate those probability values? ( nu_min = 0, 0.1, 0.2 )

 elkka 04-19-2012 04:10 AM

Thank you, Professor.

This how I calculate the probabilities. Let - the number of heads for , and let be the number of heads in i-th experiment (out of 1000). Then, as Professor has shown previously,

Now, . Therefore,

Next,

Next,

The rest can be calculated directly too, but they are essenctially 0 for the purpose of calculating the mean.

 rohanag 04-19-2012 08:44 AM

thank you for the detailed explanation.

 nyxee 07-13-2012 05:55 PM

the answer is very clear but how do we know when to use this not.

to clarify my question, if say P(ten heads)=p and P(not ten heads)=q (=1-p). why does using (p^1000) give the wrong answer?

 htlin 07-15-2012 04:56 AM

Quote:
 Originally Posted by nyxee (Post 3399) the answer is very clear but how do we know when to use this not. to clarify my question, if say P(ten heads)=p and P(not ten heads)=q (=1-p). why does using (p^1000) give the wrong answer?
means the probability of getting ten heads in each of the independent random trials. Is that the event you are interested in? ;)

 coolguy 09-02-2012 08:05 AM

Quote:
 Originally Posted by yaser (Post 1288) The probability of getting 10 heads for one coin is (10 times) which is aprroximately . Therefore, the probability of not getting 10 heads for one coin is approximately . This means that the probability of not getting 10 heads for any of 1000 coins is this number multiplied by itself 1000 times, once for every coin. This probability is therefore . This is approximately since . Numerically, . Therefore, the probability of this not happening, namely that at least one coin of the 1000 coins will give 10 heads, is 1 minus that. This gives us the answer of approximately 0.63 or 63% that I mentioned in the lecture.
WOW!Thanks for the detailed explanation!It was over my head for a long time.

 weehong 08-28-2013 03:32 AM

Quote:
 Originally Posted by yaser (Post 1288) The probability of getting 10 heads for one coin is (10 times) which is aprroximately . Therefore, the probability of not getting 10 heads for one coin is approximately . This means that the probability of not getting 10 heads for any of 1000 coins is this number multiplied by itself 1000 times, once for every coin. This probability is therefore . This is approximately since . Numerically, . Therefore, the probability of this not happening, namely that at least one coin of the 1000 coins will give 10 heads, is 1 minus that. This gives us the answer of approximately 0.63 or 63% that I mentioned in the lecture.
I captured the above idea, however I am confused why the following approach gives wrong answer:

Probability of all heads in 10 flips by one coin: p = 0.5^10
Probability of all heads in 10 flips by any of 1000 coin: 1000*p = 97.7%

In above I assumed there were 1000 10-flips.

 yaser 08-30-2013 02:02 AM

Quote:
 Originally Posted by weehong (Post 11459) I captured the above idea, however I am confused why the following approach gives wrong answer: Probability of all heads in 10 flips by one coin: p = 0.5^10 Probability of all heads in 10 flips by any of 1000 coin: 1000*p = 97.7% In above I assumed there were 1000 10-flips.
In order to multiply the probabilities in the "all" case, you need the events to be independent, and we have that for the coin flips. In order to add the probabilities in the "any of" case, you need the events to be disjoint, i.e., they cannot simultaneously occur. The first coin giving 10 heads is not disjoint from the second coin giving 10 heads, so when you add their probabilities you are double counting the overlap which is "both coins giving 10 heads each."

 weehong 09-01-2013 12:41 AM