Opensource Implementation of the Alias Method [closed]
-
02-07-2019 - |
Question
I am doing a project at the moment, and in the interest of code reuse, I went looking for a library that can perform some probabilistic accept/reject of an item:
i.e., there are three people (a, b c), and each of them have a probability P{i} of getting an item, where p{a} denotes the probability of a. These probabilities are calculated at run time, and cannot be hardcoded.
What I wanted to do is to generate one random number (for an item), and calculate who gets that item based on their probability of getting it. The alias method (http://books.google.com/books?pg=PA133&dq=alias+method+walker&ei=D4ORR8ncFYuWtgOslpVE&sig=TjEThBUa4odbGJmjyF4daF1AKF4&id=ERSSDBDcYOIC&output=html) outlined here explained how, but I wanted to see if there is a ready made implementation so I wouldn't have to write it up.
Solution
Would something like this do? Put all p{i}'s in the array, function will return an index to the person who gets the item. Executes in O(n).
public int selectPerson(float[] probabilies, Random r) {
float t = r.nextFloat();
float p = 0.0f;
for (int i = 0; i < probabilies.length; i++) {
p += probabilies[i];
if (t < p) {
return i;
}
}
// We should not end up here if probabilities are normalized properly (sum up to one)
return probabilies.length - 1;
}
EDIT: I haven't really tested this. My point was that the function you described is not very complicated (if I understood what you meant correctly, that is), and you shouldn't need to download a library to solve this.
OTHER TIPS
Here is a Ruby implementation: https://github.com/cantino/walker_method
i just tested out the method above - its not perfect, but i guess for my purposes, it ought to be enough. (code in groovy, pasted into a unit test...)
void test() {
for (int i = 0; i < 10; i++) {
once()
}
}
private def once() {
def double[] probs = [1 / 11, 2 / 11, 3 / 11, 1 / 11, 2 / 11, 2 / 11]
def int[] whoCounts = new int[probs.length]
def Random r = new Random()
def int who
int TIMES = 1000000
for (int i = 0; i < TIMES; i++) {
who = selectPerson(probs, r.nextDouble())
whoCounts[who]++
}
for (int j = 0; j < probs.length; j++) {
System.out.printf(" %10f ", (probs[j] - (whoCounts[j] / TIMES)))
}
println ""
}
public int selectPerson(double[] probabilies, double r) {
double t = r
double p = 0.0f;
for (int i = 0; i < probabilies.length; i++) {
p += probabilies[i];
if (t < p) {
return i;
}
}
return probabilies.length - 1;
}
outputs: the difference betweenn the probability, and the actual count/total
obtained over ten 1,000,000 runs:
-0.000009 0.000027 0.000149 -0.000125 0.000371 -0.000414
-0.000212 -0.000346 -0.000396 0.000013 0.000808 0.000132
0.000326 0.000231 -0.000113 0.000040 -0.000071 -0.000414
0.000236 0.000390 -0.000733 -0.000368 0.000086 0.000388
-0.000202 -0.000473 -0.000250 0.000101 -0.000140 0.000963
0.000076 0.000487 -0.000106 -0.000044 0.000095 -0.000509
0.000295 0.000117 -0.000545 -0.000112 -0.000062 0.000306
-0.000584 0.000651 0.000191 0.000280 -0.000358 -0.000181
-0.000334 -0.000043 0.000484 -0.000156 0.000420 -0.000372