A Julia package for Stochastic Dynamic Programming
Author odow
25 Stars
Updated Last
1 Year Ago
Started In
September 2016


A Julia package for discrete stochastic dynamic programming.

Build Status Coverage
Build Status Codecov branch


This package is not yet registered. Until it is, things may change. It is perfectly usable but should not be considered stable.

You can install it by typing

julia> Pkg.add(url="https://github.com/odow/DynamicProgramming.jl")

Initialise Model

A SDP model is stored in the SDPModel type. A SDPModel can be intialised using the following SDPModel() do ... end block structure:

m = SDPModel(
        stages = 1.   # Int
        sense  = :Min # Symbol (:Min or :Max)
    ) do sp, t

        ... problem definition ...


sp can be given any name but refers to the stage problem. t can also be given any name, but refers to an index that runs from 1 to T where T is the number of stages.

Inside the SDPModel definition, we define our subproblems. We first need to add some state variables, some control (or action) variables, and some noise (or stochastic) variables.

Initialise Variables

States can be added with the following macro:

@states(sp, begin
    x in linspace(0, 1, 10)

This creates a state variable x that is discretised into the set linspace(0, 1, 10). Note that currently, all state dimensions get converted into Float64 representations. The discretisation should be any type that can be converted to a Vector{Float64} type.

Controls can be added with the @controls macro that has similar syntax. However there is less restriction on the type. The discretisation should just be an iterable subtype of AbstractVector.

Noise (or stochastic variables) can be added with the @noises macro:

@noises(sp, begin
    u in DiscreteDistribution([1,2,3], [0.5, 0.25, 0.25])
    v in 1:10

In contrast to the other two macros, there is a slight subtlety. The discretisations can either be subtypes of AbstractVector (in which case their realisations are assumed to be uniformly sampled), or a DiscreteDistribution.

The DiscreteDistribution constructor is DiscreteDistribution(observations::AbstractVector, probability::AbstractVector). This realisations observations are sampled with probability probability.

If more than one noise is defined, then the multiple noises are assumed to be independent.


You must provide a function that takes four inputs.

function foo(states_out, states, controls, noises)

However, we prefer the anonymous function syntax:

dynamics!(sp) do y, x, u, w
        ... definitions ...

This anonymous function must take the current state x, a control u and a noise w and update the new state y.

You can refer to model variables using the [] indexing operator. For example, if we defined a state variable quantity, we could refer it as x[quantity].

By thinking about variable scopes it is possible to encapsulate all the necessary data into this syntax.

Important: the function should return a single Float64 value corresponding to the cost (or profit) accrued in the current stage.

Terminal Objective

The terminal objective function takes as input a vector of the final state at the end of the finite time horizon. It returns a single Float64 value corresponding to the cost (or profit) of ending in that state.

terminalobjective!(sp) do x
        ... definitions ...


The constraints function takes as input vectors for the initial state, control and noise. It should return a single Bool value indicating if the state, control, noise combination is feasible. Typically this can be implemented by a chained series of boolean comparisons.

constraints!(sp) do x, u, w
        ... definitions ...


The solve function takes as input the initalised SDPModel object, as well as two keyword arguments.

The realisation must be one of WaitAndSee, HereAndNow or ExpectedValue.

The WaitAndSee model observes the noise before choosing the optimal control. The HereAndNow model chooses the best control before observing the noise. The ExpectedValue model substitutes the noise for the expected value of each independent noise.

The riskmeasure is a nested λE[x] + (1-λ)CVaRᵦ[x].

    riskmeasure=NestedCVaR(beta=0.5, lambda=0.5)


Once a SDPModel has been solved, it is possible to simulate the performance of the policy using the function simulate(m::SDPModel, N::Int; kwargs...). m is the solved SDPModel to be simulated. N is the number of realisations to perform. Initial values for the state variables are given via the keyword arguments.

For example:

results = simulate(m,
    contracts  = 0,
    price      = 4.5,
    production = 0.


It is possible to create an interactive visualisation of the simulated policy with the @visualise macro. The following keywords should be wrapped with parentheses.

  • "cumulative" = false Plot the cumulation of the variable over stages
  • "title" = "" Plot title
  • "xlabel" = "Stages" Label for x axis
  • "ylabel" = "" Label for y axis
  • "interpolate" = "linear" D3.js interpolation method to use. See the D3 wiki for more.

The following example gives an example of possible syntax:

@visualise(results, stage, replication, begin
    results[:Current][stage][replication],  (title="Accumulated Profit", ylabel="Accumulated Profit (\$)", cumulative=true)
    results[:x][stage][replication],    (title="Value of a State", ylabel="Level")
    results[:u][stage][replication],    (title="Value of a Control")
    results[:w][stage][replication],    (title="Value of a Noise", interpolate="step")