MATLAB Answers

0

Random sample without replacement

Hi,all,
Does anybody know how to do random sample without replacement? The randsample function in matlab only supports sampling with replacement.
I made codes on my own, and it is really weird sometimes it works, but sometimes it shows error (Error using ==> randsample at 94 W must have length equal to N.):
function C=randsample_WithoutReplacement(m,n,A1,A2)
%A1:population
%A2:probability
B=zeros(m,1);
C=zeros(n,m);
s=transpose(1:1:length(A1));
ut=0;
loc=0;
A=A2;
for j=1:n
A=A2;
s=transpose(1:1:length(A1));
for i=1:m
B(i)=randsample(s,1,true,A);
[ut, loc] = ismember(B(i), s);
s(loc)=[];
A(loc)=[];
end
for i=1:m
C(j,i)=A1(B(i));
end
end

  3 Comments

What version of MATLAB do you have? In R2011a, randsample supports sampling with or without replacement.
Oh - you mean it doesn't support weighted sampling without replacement.
I can't reproduce your error.

Sign in to comment.

3 Answers

Answer by Andrew Newell
on 31 Jan 2012
 Accepted Answer

Here is a recursive function that will return one row of your matrix C:
function y = randsampleWithoutReplacement(population,k,w)
y = [];
if ~isempty(population) && k > 0
n = length(population);
ii = randsample(1:n,1,true,w);
newpop = setdiff(1:n,ii);
y = [population(ii) randsampleWithoutReplacement(population(newpop),k-1,w(newpop))];
end
and here is an example of a call:
y = randsampleWithoutReplacement(1:100,20,ones(100,1)/100)
EDIT: And here is one that is closer to your version:
function C=randsample_WithoutReplacement(A1,k,A2)
%A1:population
%A2:probability
C=zeros(1,k);
A=A2;
n = length(A1);
for i=1:k
loc=randsample(n-i+1,1,true,A);
A(loc)=[];
C(i)=A1(loc);
end
However, I have designed it to act more like randsample.

  1 Comment

Thanks, I have developed a stable version, and yours seem to work also. I found maybe the error comes from how I called the function.

Sign in to comment.


Answer by Peter Perkins
on 31 Jan 2012

As Andrew pointed out, randsample absolutely does do sampling without replacement, just not with weights. It looks like that's what you're asking for.
If you have access to R2011b, you can use the new datasample function in the Statistics Toolbox (a replacement for randsample, though randsample continues to work) for sampling with and without replacement, weighted or unweighted:

  1 Comment

Thanks, I'll also look into the R2011b new function.

Sign in to comment.


Answer by Derek O'Connor on 31 Jan 2012

If you don't have access to R2011b and randsample, then the function below is reasonably fast on my Dell Precision 690, 2.33GHz, 16GB ram, Windows 7 Professional, Matlab R2008b 64-bit.
It uses a rejection loop to call DiscITBS, which generates a single sample from a discrete distribution by doing a binary search on the CDF, which, by definition, is sorted in ascending order.
Membership in S is tested by the byte-array member. This is a bit expensive (of memory) but is fast and simple. If you have lots of memory, then use it.
The expected value of the running time is Ns*Ew*log(Np), where Ew = E(nw) is the expected number of trips around the rejection loop.
If Np is small then it doesn't matter what method you use. If Np = 10^6, and Ns < 0.25*Np then this method is quite fast because Ew will be small and log(Np) of binary search takes care of the large Np.
For example:
with Np = 10^6 and Ns = 10^3, nw = 3 and t = 0.05 secs.
with Np = 10^6 and Ns = 10^4, nw = 129 and t = 0.23 secs.
% -------------------------------------------------------------
function [S,nw] = DiscSampRej(x,p,Ns);
% -------------------------------------------------------------
% Generate a random sample of size Ns from x(1:Np) with prob
% p(1:Np), without replacement. Derek O'Connor 31 Jan 2012
% -------------------------------------------------------------
S = zeros(1,Ns);
Np = length(x);
member(1:Np) = false;
cdf = cumsum(p);
nw=0;
for k = 1:Ns
idx = DiscITBS(cdf);
while member(idx)
idx = DiscITBS(cdf);nw=nw+1;
end
S(k) = x(idx);
member(idx) = true;
end % function
% -------------------------------------------------------------
function idx = DiscITBS(cdf);
% -------------------------------------------------------------
% Uses the discrete Inverse Transform method with Binary Search
% This greatly reduces the number of iterations of the while-loop
% Time Complexity: O(log n)
u = rand;
L = 1; H = length(cdf);
while L <= H
m = floor(L/2+H/2);
if u < cdf(m)
H = m-1;
else
L = m+1;
end
end
idx = m;
% end function

  0 Comments

Sign in to comment.