Elixir weighted random pick library optimised for quick take_one
and take_n
operations.
take_one\1
and take_n\2
behaviours are compatible with Enum.random\1
and Enum.take_random\2
and can almost be drop-in replacements, except that WeightedRandom
accepts elements in format {element :: any(), weight :: number()}
.
The package can be installed
by adding better_weighted_random
to your list of dependencies in mix.exs
:
def deps do
[
{:better_weighted_random, "~> 0.1.0"}
]
end
WeightedRandom.take_one\1
returns one random element from input, with probability based on weights.
iex> :rand.seed(:exsss, {100, 101, 102})
iex> WeightedRandom.take_one([{:'1', 0.5}, {:'2', 1.0}, {:'3', 2.0}])
:"3"
WeightedRandom.take_n\1
returns n random elements from input, with probability based on weights. Similar to Enum.take_random\2
element can not be picked twice.
iex> :rand.seed(:exsss, {100, 101, 102})
iex> WeightedRandom.take_n([{:'1', 0.5}, {:'2', 1.0}, {:'3', 2.0}], 2)
[:"3", :"1"]
Probability of picking an element directly depends on its weight. Probability = element_weight / sum(all_weights)
iex> :rand.seed(:exsss, {100, 101, 102})
iex> Enum.map(1..1000, fn _ -> WeightedRandom.take_one([{:'1', 1.0}, {:'2', 2.0}, {:'3', 1.0}]) end) |> Enum.frequencies()
%{
"1": 231,
"2": 522,
"3": 247
}
You can create searcher with WeightedRandom.create_searcher\1
for optimised lookups with take_one\1
and take_n\2
.
iex> searcher = WeightedRandom.create_searcher([{:'1', 1.0}, {:'2', 2.0}, {:'3', 1.0}])
iex> :rand.seed(:exsss, {100, 101, 102})
iex> WeightedRandom.take_one(searcher)
:"3"
iex> rand.seed(:exsss, {100, 101, 102})
iex> WeightedRandom.take_n(searcher, 2)
[:"3", :"1"]
Using searcher is quicker if you do repetative calls.
elements = Enum.map(1..10000, fn i -> {:"el#{i}", :rand.uniform()} end)
searcher = WeightedRandom.create_searcher(elements)
Name ips average deviation median 99th %
take_one with searcher 787.69 K 0.00127 ms ±2423.14% 0.00109 ms 0.00180 ms
take_one with list 0.137 K 7.29 ms ±17.57% 7.03 ms 11.42 ms
Comparison:
take_one with searcher 787.69 K
take_one with list 0.137 K - 5740.48x slower +7.29 ms
Name ips average deviation median 99th %
take_n 100 with searcher 7.26 K 0.138 ms ±30.77% 0.124 ms 0.32 ms
take_n 100 with list 0.106 K 9.43 ms ±12.97% 8.96 ms 12.29 ms
Comparison:
take_n 100 with searcher 7.26 K
take_n 100 with list 0.106 K - 68.46x slower +9.30 ms
Docs are availiable at https://hexdocs.pm/better_weighted_random