Iterators with message passing and feedback loops




Iterators combine to a tree of iterators, but dynamic iterators combine to a network of interacting entities.

Dynamic iterators subtype <:DynamicIterator. They extend the iteration protocol and define

    dyniterate(iter, somemessage(state))


    dyniterate(iter, othermessage(state), arg)

where message wraps a state or other relevant information. For example the definition

struct Start{T} <: Message
dyniterate(iter, Start(value))

communicates that iter should start at value (if this is implemented). This is similar to iterate(iter) communicating that iter should start at a predefined value. In fact a fallback

dyniterate(iter, ::Nothing) = iterate(iter)

is in place.

Some messages make the iterator accept a third argument. A simple example using bind to bind an iterator to an iterator using the three-argument form of dyniterate:

using DynamicIterators
import DynamicIterators: dyniterate

struct Summed <: DynamicIterator

function dyniterate(::Summed, ::Nothing, y)
    y, y

function dyniterate(::Summed, i, y)
    i + y, i + y

@show collect(bind(1:5, Summed()))

A more in-depth example showing the power of the approach is https://github.com/mschauer/DynamicIterators.jl/blob/master/example/ressourcemanagement.jl, showing how to extend the iterator protocol to allow resource management (e.g. closing of files of child iterators) at the end of iteration of the parent.

A preliminary list of supported messages:

Message (and third argument) Meaning
state or State(state) ordinary iteration
Start(noting) start the iterator at its default
Start(x) start the iterate from the state corresponding to value x
Value(x, state) continue to iterate from the state corresponding to iterate x
NextKey(state, nextkey) advance an iterator over pairs of key=>values to nextkey
Steps(state, n) advance the iterator n steps or possibly rewind if n negative
Control(state), control control term as in the Kalman filter provided as third argument to dyniterate⋆
Sample(state[,rng]) sample from iterates⋆
NextKeys(state), key advance iterator to the keys provided as third argument to dyniterate⋆

⋆persistent messages: dyniterate returns a state again wrapped by the message

Evolution: Evolution-type dynamic iterators

Typically, the state of an iterator is opaque. But for some iterators the iterates are the states:

julia> value, state = iterate('A':'Z')
('A', 'A')

julia> value, state = iterate('A':'Z', 'X')
('Y', 'Y')

This means that the states/iterates of an iterator can be modified in a transparent way. This allows iterators not only to depend on each other, but to interact.

DynamicIterators.jl embeds a constrained iterator protocol for iterators subtyping <:Evolution, which define

evolve(iterator, x) -> y
dub(x) = x === nothing ? nothing : (x,x)
iterate(iterator::Evolution, x) = dub(evolve(iterator, x))

which guarantees value == state and introduces a powerful set of combinators for such iterators.


As a simple example take a Metropolis-Hastings chain

It can be described as a simple Evolution.

function evolve(MH::MetropolisHastings, (t,x)::Pair)
    P = MH.P
    Q = MH.proposal(x)
    xᵒ = rand(Q)
    Qᵒ = MH.proposal(xᵒ)
    if log(rand(MH.rng)) < MH.logpdf(P, xᵒ) - MH.logpdf(P, x) + MH.logpdf(Qᵒ, x) - MH.logpdf(Q, xᵒ)
        x = xᵒ
    (t+1 => x)

The following example shows that the Mixture iterator combinator can be used to combine two Metropolis-Hastings chains into a component wise MetropolisHastings sampler:

using DynamicIterators
using Distributions

D = MvNormal([1.0, 0.5], [1.0 0.5; 0.5 1.5] )
struct Move{T}
m1(x) = Move(x, 0.1, 1)
m2(x) = Move(x, 0.1, 2)
Base.rand(M::Move) = M.x + M.σ*randn()*[M.i-1, 2-M.i]
Distributions.logpdf(M::Move, x) = logpdf(Normal(M.x[M.i], M.σ), x[M.i])
MH1 = MetropolisHastings(D, m1, logpdf)
MH2 = MetropolisHastings(D, m2, logpdf)

I = Evolve(i->rand(1:2))

MH = mixture(I, (MH1, MH2))

X = values(trace(MH, 1=>(1, [0.0, 0.0]), endtime(2000)))


Lifting time


evolve(E, (i, x)::Pair) = i + 1 => evolve(E, x)

constitutes a "lifting" of discrete time. This corresponds to enumerating the iterates of an evolution x = f(x) as (1 => x1, 2 => x2, ...).

DynamicIterators control keywords treat Pairs as pair of key and value in concordance with the package Trajectories and somewhat in line with Julia's general convention.


Controlled Dynamic Iterators


To illustrates the range of this I have picked some examples of very diverse nature.

First Commit


Last Touched

2 months ago


66 commits

Used By: