Consensus.jl is a lightweight, gradient-free, stochastic optimisation package for Julia. It uses Consensus-Based Optimisation (CBO), a flavour of Particle Swarm Optimisation (PSO) first introduced by R. Pinnau, C. Totzeck, O. Tse, and S. Martin (2017). This is a method of global optimisation particularly suited for rough functions, where gradient descent would fail. It is also useful for optimisation in higher dimensions.

This package was created and is developed by Dr Rafael Bailo.


The basic command of the library is minimise(f, x0), where f is the function you want to minimise, and x0 is an initial guess. It returns an approximation of the point x that minimises f.

You have two options to define the objective function:

  • x is of type Real, and f is defined as f(x::Real) = ....
  • x is of type AbstractVector{<:Real}, and f is defined as f(x::AbstractVector{<:Real}) = ....

A trivial example

We can demonstrate the functionality of the library by minimising the function $f(x)=x^2$. If you suspect the minimiser is near $x=1$, you can simply run

using Consensus
f(x) = x^2;
x0 = 1;
x = minimise(f, x0)

to obtain

julia> x
1-element Vector{Float64}:

Your x may vary, since the method is stochastic. The answer should be close, but not exactly equal, to zero.

Behind the scenes, Consensus.jl is running the algorithm using N = 50 particles per realisation. It runs the M = 100 realisations, and returns the averaged result. If you want to parallelise these runs, simply start julia with multiple threads, e.g.:

$ julia --threads 4

Consensus.jl will then automatically parallelise the optimisation. This is thanks to the functionality of StochasticDiffEq.jl, which is used under the hood to implement the algorithm.

Advanced options

There are several parameters that can be customised. The most important are:

  • N: the number of particles per realisation.
  • M: the number of realisations, whose results are averaged in the end.
  • T: the run time of each realisation. The longer this is, the better the results, but the longer you have to wait for them.
  • Δt: the discretisation step of the realisations. Smaller is more accurate, but slower. If the optimisation fails (returns Inf or NaN), making this smaller is likely to help.
  • R: the radius of the initial sampling area, which is centred around your intiial guess x0.
  • α: the exponential weight. The higher this is, the better the results, but you might need to decrease Δt if α is too large.

We can run the previous example with custom parameters by calling

julia> x2 = minimise(f, x0, N = 30, M = 100, T = 10, Δt = 0.5, R = 2, α = 500)
1-element Vector{Float64}:

For the other parameters, please refer to the paper of R. Pinnau, C. Totzeck, O. Tse, and S. Martin (2017). You can see the default values of the parameters by evaluating Consensus.DEFAULT_OPTIONS.

Non-trivial examples

Since CBO is not a gradient method, it will perform well on rough functions. Consensus.jl implements two well-known test cases in any number of dimensions:

We can minimise the Ackley function in two dimensions, starting near the point $x=(1,1)$, by running

julia> x3 = minimise(AckleyFunction, [1, 1])
2-element Vector{Float64}:

We can also minimise the Rastrigin function in five dimensions, starting at a random point, with more realisations, and with a larger radius, by running

julia> x4 = minimise(RastriginFunction, rand(5), M = 200, R = 5)
5-element Vector{Float64}:

Auxiliary commands

There is a maximise(f, x0) method, which simply minimises the function g(x) = -f(x). Also, if you're that way inclined, you can call minimize(f, x0) and maximize(f, x0), in the American spelling.