n******0
发帖数: 298 | 1 我有10个subjects,每个subject有一个binary variable Z1 和一个normally
distributed continuous variable Z2。 Z2的distribution在Z1=0和Z1=1时是不一样
的（比如不同的mean)，subjects是互相独立的，我需要generate这样的20
dimensional space,以求某种expectation,那位大牛知道有没有和用什么样的sequence
来做？ |
j******n
发帖数: 271 | 2 阿弥托佛 这个太易了。直接抽样不行么？
sequence
【在 n******0 的大作中提到】
: 我有10个subjects,每个subject有一个binary variable Z1 和一个normally
: distributed continuous variable Z2。 Z2的distribution在Z1=0和Z1=1时是不一样
: 的（比如不同的mean)，subjects是互相独立的，我需要generate这样的20
: dimensional space,以求某种expectation,那位大牛知道有没有和用什么样的sequence
: 来做？
|
n******0
发帖数: 298 | 3 直接抽样应该是monte carlo integration吧？我想找一个convergence更快的方法。
【在 j******n 的大作中提到】
: 阿弥托佛 这个太易了。直接抽样不行么？
:
: sequence
|
s***e
发帖数: 267 | 4 Yes quasi-MC can reduce the variance faster, and it is not a MC method
although the name contains MC.
I think for one dimension you can split the quantiles and take equal spaces.
For higher dimension you may consider the hilbert curve method to fill the
space:
http://en.wikipedia.org/wiki/Hilbert_curve
For your problem with dimension = 20, the computation could be expansive.
sequence
【在 n******0 的大作中提到】
: 我有10个subjects,每个subject有一个binary variable Z1 和一个normally
: distributed continuous variable Z2。 Z2的distribution在Z1=0和Z1=1时是不一样
: 的（比如不同的mean)，subjects是互相独立的，我需要generate这样的20
: dimensional space,以求某种expectation,那位大牛知道有没有和用什么样的sequence
: 来做？
|
n******0
发帖数: 298 | 5 The number of points in the space should not increase with the number of
dimensions. This is the exact reason I am considering quasi-monte carlo. The
question is how to generate these points to deal with my specific problem
with mixed binary and normal distributions with normal dist depending on the
value of the binary variable.
spaces.
the
【在 s***e 的大作中提到】
: Yes quasi-MC can reduce the variance faster, and it is not a MC method
: although the name contains MC.
: I think for one dimension you can split the quantiles and take equal spaces.
: For higher dimension you may consider the hilbert curve method to fill the
: space:
: http://en.wikipedia.org/wiki/Hilbert_curve
: For your problem with dimension = 20, the computation could be expansive.
:
: sequence
|
s***e
发帖数: 267 | 6 In that case, one possible approach is to fill a lower dimensional space,
say, and randomly sample the rest.
I think the higher the dimension, the smaller the advantage of quasi-MC over
MC. The variance of MC is dimension independent (although not good enough
for many applications).
The
the
【在 n******0 的大作中提到】
: The number of points in the space should not increase with the number of
: dimensions. This is the exact reason I am considering quasi-monte carlo. The
: question is how to generate these points to deal with my specific problem
: with mixed binary and normal distributions with normal dist depending on the
: value of the binary variable.
:
: spaces.
: the
|
