fy0 / weightedrand

:balance_scale: Fast weighted random selection for Go

weightedrand ⚖️

Fast weighted random selection for Go.

Randomly selects an element from some kind of list, where the chances of each element to be selected are not equal, but rather defined by relative "weights" (or probabilities). This is called weighted random selection.

Usage

import (
    /* ...snip... */
    wr "github.com/mroth/weightedrand"
)

func main() {
    rand.Seed(time.Now().UTC().UnixNano()) // always seed random!

    chooser, _ := wr.NewChooser(
        wr.Choice{Item: "🍒", Weight: 0},
        wr.Choice{Item: "🍋", Weight: 1},
        wr.Choice{Item: "🍊", Weight: 1},
        wr.Choice{Item: "🍉", Weight: 3},
        wr.Choice{Item: "🥑", Weight: 5},
    )
    /* The following will print 🍋 and 🍊 with 0.1 probability, 🍉 with 0.3
    probability, and 🥑 with 0.5 probability. 🍒 will never be printed. (Note
    the weights don't have to add up to 10, that was just done here to make the
    example easier to read.) */
    result := chooser.Pick().(string)
    fmt.Println(result)
}

Performance

The existing Go library that has a comparable implementation of this is github.com/jmcvetta/randutil, which optimizes for the single operation case. In contrast, this library creates a presorted cache optimized for binary search, allowing repeated selections from the same set to be significantly faster, especially for large data sets.

Comparison of this library versus randutil.ChooseWeighted on my workstation. For repeated samplings from large collections, weightedrand will be much quicker:

Num choices	randutil	weightedrand	weightedrand -cpu=8*
10	201 ns/op	38 ns/op	2.9 ns/op
100	267 ns/op	51 ns/op	4.1 ns/op
1,000	1012 ns/op	67 ns/op	5.4 ns/op
10,000	8683 ns/op	83 ns/op	6.9 ns/op
100,000	123500 ns/op	105 ns/op	12.0 ns/op
1,000,000	2399614 ns/op	218 ns/op	17.2 ns/op
10,000,000	26804440 ns/op	432 ns/op	35.1 ns/op

*: Since v0.3.0 weightedrand can efficiently utilize a single Chooser across multiple CPU cores in parallel, making it even faster in overall throughput. See PR#2 for details. Informal benchmarks conducted on an Intel Xeon W-2140B CPU (8 core @ 3.2GHz, hyperthreading enabled).

Don't be mislead by these numbers into thinking weightedrand is always the right choice! If you are only picking from the same distribution once, randutil will be faster. weightedrand optimizes for repeated calls at the expense of some initialization time and memory storage.

Caveats

Note this library utilizes math/rand instead of crypto/rand, as it is optimized for performance, and is not intended to be used for cryptographically secure requirements.

Credits

To better understand the algorithm used in this library (as well as the one used in randutil) check out this great blog post: Weighted random generation in Python.