Solid tutorial, thanks so much! My one suggestion for making it even better would be to tweak the details of the Likehood Weighting example a bit (starts at 17:50). Basically, I was a bit confused for a while how one could get different weight values from the samples, because in this example the samples will indeed always return the same results since they are only controlled by evidence variables. I think changing the example to 'find P(-d | -b, -c)' would work (swap out a for b, to allow a to be randomized to change which area P(b | a) and P(c | a) pull from).
Hi Aram, yes, likelihood weighting is an instantiation of importance sampling. In general in importance sampling you were interested in getting samples from a distribution Q1, but unfortunately you don't know how to sample efficiently from Q1 so you instead sample from another distribution Q2 that is easier to sample from --- and then you reweight your samples by the ratio Q1/Q2. In likelihood weighting: Q1 = P(unobserved variables | observed variables), Q2 = defined by the sampling process.
I too have the same question. GIven the evidence (-a,-b), we don't sample 'a' and 'b' from the distribution and we sample only c and d. But still there are samples with +a,+b / +a,-b / -a,+b combinations in 20:58!! how are the combined weights calculated? Thanks in advance. Thanks for the clear and simple explanation given.
@@sathblr In general, when you do Likelihood weighting you have to weigh each sample by the likelihood of that generated sample with the evidence variables ("how much that sample makes sense, given the evidence of your query") This means the total weight is the product of the weights for each evidence variable (remember: evidence variables are always fixed by the query): conditional probability of the evidence given its parents. In this particular case, though, the evidence variables are 'a' and 'b': - 'a' doesn't have parents. - 'b' has only 'a' as a parent. For this reason, the weight of each sample is given by the following: P(A) * P(B|A), where B and A are actual settings of the variables for the current sample. Example: In the last sample (21:01) we have "+a , +b , -c , -d" --> weight = P(+a) * P(+b | +a) = 0.8 * 0.8 = 0.64 For this reason, I believe the weights provided by the Teacher are just for the sake of explaining the concepts. Please, correct me if I'm wrong. I wish this helped.
Even after 11 years it's one of the best videos. Thank you
Best and simplest thing I've found on the web so far about sampling on Bayes' nets!
Excellent explanation. I appreciate it so much. Now i finally understand this after so much struggling looking at some other resources. Thank you!
Dear Mr. Abbeel, thank you for your explanation! It has been 10 years and this tutorial still rocks :)
At 24:00, you added the last sample that had +d even though you were conditioning on -d. Just a small error.
greatest explanation on an example. Thank you professor Pieter!
Thank you sir!!!
Good video, it's actually very simple when you see it in practice and not just as an abstract set of equations.
Solid tutorial, thanks so much! My one suggestion for making it even better would be to tweak the details of the Likehood Weighting example a bit (starts at 17:50). Basically, I was a bit confused for a while how one could get different weight values from the samples, because in this example the samples will indeed always return the same results since they are only controlled by evidence variables. I think changing the example to 'find P(-d | -b, -c)' would work (swap out a for b, to allow a to be randomized to change which area P(b | a) and P(c | a) pull from).
Waaay better than anything I've seen so far. Much respect!
Thank you so much you saved my life
This is the better video and the better teacher!
Excellent description of Bayesian network sampling!
thank you, and glad it's helpful!
Thank you so so so much
My professor could learn a thing from you
This video is so good and clear!
Thank you very much for this! I wish you get everything you want in life
what the hell? are you from kz?
@@figuraclass444 yes i am
great tutorial Mr.Abbeel :)❤
Great explanation !!!
In the last part you fixed B=-b and a=-a but the samples you showed did not reflect this. That part is extremely wrong and needs rectification.
at 9:04 does it really map to +c? 0.04 seems like it would map to -c.
The best tutorial about sampling. Thanks :)
Very clear explaination. Awesome video. thanks a lot!
Excellent video
Thank you for this simple lesson!
Thank you sooo much. You just saved my exam
Thank you professor, I really became a fan of your teaching style. One question: is likelihood weighting is the same as Gibbs sampling?
Hi Aram, yes, likelihood weighting is an instantiation of importance sampling. In general in importance sampling you were interested in getting samples from a distribution Q1, but unfortunately you don't know how to sample efficiently from Q1 so you instead sample from another distribution Q2 that is easier to sample from --- and then you reweight your samples by the ratio Q1/Q2. In likelihood weighting: Q1 = P(unobserved variables | observed variables), Q2 = defined by the sampling process.
best explanation of all !
Great explanation sir, thank you very much!
Very helpful! Thank you very much
Best explanation!
Still best video until 2019
Dear Dr., in time 10:06, is the answering probabilistic queries uses for crating condition probability in Genie software
Excellent.. But how to do sampling when the variables are continuous...
While doing likelihood weighting for the last question, while collecting samples how you would proceed? Given there are multiple conditional queries.
You can directly add up the weights containing those (evidence + queries) upon the summation over the weights of the evidence
Hi, Where did you get the examples?
Very clear. Thank you.
excellent
Excuse me, I meant importance sampling.is likelihood weighting same as importance sampling?
Bless
Best video
How did you find the weights of each sample in 20:58 ? thanks in advance
I too have the same question. GIven the evidence (-a,-b), we don't sample 'a' and 'b' from the distribution and we sample only c and d. But still there are samples with +a,+b / +a,-b / -a,+b combinations in 20:58!! how are the combined weights calculated? Thanks in advance. Thanks for the clear and simple explanation given.
@@sathblr In general, when you do Likelihood weighting you have to weigh each sample by the likelihood of that generated sample with the evidence variables ("how much that sample makes sense, given the evidence of your query")
This means the total weight is the product of the weights for each evidence variable (remember: evidence variables are always fixed by the query): conditional probability of the evidence given its parents.
In this particular case, though, the evidence variables are 'a' and 'b':
- 'a' doesn't have parents.
- 'b' has only 'a' as a parent.
For this reason, the weight of each sample is given by the following: P(A) * P(B|A), where B and A are actual settings of the variables for the current sample.
Example:
In the last sample (21:01) we have "+a , +b , -c , -d" --> weight = P(+a) * P(+b | +a) = 0.8 * 0.8 = 0.64
For this reason, I believe the weights provided by the Teacher are just for the sake of explaining the concepts.
Please, correct me if I'm wrong. I wish this helped.
Nice explanation
God Explaination
Please make more small AI videos on different topics having good question examples.
i should be paying my college fees to this guy.
Go bears!
0:00 earrape alert
19:46 0.2 * 0.5 =/= 0.1
From Rejection Sampling onwards, the explanation is bad. It is too fast.
This is a good video, but you're talking too fast.
this video is understandable even without audio
Good, but you were too fast, and the words were hard too.
Can ya talk any fucken faster, who ya tryin to impress, a machinegun?
lmfao :'D
Dude get some space between your mouth and mic, and talk a bit slower...