Comments on: Weighted random generation in Python

By: pronostic

pronostic — Tue, 27 Jul 2010 15:02:13 +0000

Thanks for sharing this “Walker’s alias method”. Very interesting !

By: denis

denis — Mon, 14 Jun 2010 14:33:37 +0000

hi Eli,
you might look at Walker’s alias method, in Knuth Graphbase p. 392 or
http://code.activestate.com/recipes/576564-walkers-alias-method-for-random-objects-with-diffe/

and keep up the good posts.
(Care to start a Real Doc project, collecting best sw doc in various categories ?)

cheers
— denis

By: eliben

eliben — Sat, 23 Jan 2010 07:48:19 +0000

@Evan, interesting, thanks. The numpy version is a bit faster than the bisection method for very long lists, but for short lists it isn’t very good. The method Simon presents is faster in all cases

@Simon, nice variation! It’s really fast – I think I’ll even add it to the article.

By: Simon Davy

Simon Davy — Fri, 22 Jan 2010 23:47:11 +0000

Similar to a “roulette” select, it’s used a lot in genetic algorithms[1].

In the first version, I think this would be better as it needs no extra list.

choice = random() * sum(weights)
for i, w in enumerate(weights):
    choice -= w
    if choice < 0:
        return i

If you can contrive to order the weights in descending order, then this would definitely be quicker I think.

If you can prebuild the totals list, then the bisect is probably faster.

Of course, if you can guarantee the weights are ints you can build a simple list and use a single random int look up.

I think this may be covered in TAOCP – but then, what isn’t?

[1] http://en.wikipedia.org/wiki/Fitness_proportionate_selection

By: Richard Tew

Richard Tew — Fri, 22 Jan 2010 18:57:35 +0000

This should be in the ‘random’ standard library module. It’s just one of those things that you end up writing in the course of the development of a code base. I used it all over the place in EVE Online, from NPC spawning to loot generation.

By: Evan Friis

Evan Friis — Fri, 22 Jan 2010 17:18:14 +0000

There is also a pretty handy way to do this with numpy:

import numpy as np

def weighted_choice(weights):
    totals = np.cumsum(weights)
    norm = totals[-1]
    throw = np.random.rand()*norm
    return np.searchsorted(totals, throw)