Minfinder

The MinFinder algorithm solves those problems where you need to find all the minima for a differentiable function inside a bounded domain. It launches many local optimizations from a set of random starting points in the search domain, after some preselection of the points and until a stopping criteria is hit. The local searches use the Fminbox method from the Optim.jl package.

This package originated from a pull request for the Optim.jl package but now simply extends that package.

I have some ideas for some extra features, but do let me know in the issues if you have more. For example:

• use low-discrepancy samples for the starting points from the Sobol.jl package
• implement 2 more stopping rules from the MinFinder 2.0 paper, as well as the extra sample checking rule without derivatives.

Have a good look at the Fminbox section of Optim.jl, because you need to pass your function in a Optim.DifferentiableFunction type.

As an example, consider the Six Hump Camel Back function with 6 minima inside [-5, 5]²:

camel_f(x) = 4x[1]^2 - 2.1x[1]^4 + 1/3*x[1]^6 + x[1]*x[2] - 4x[2]^2 + 4x[2]^4

function camel_g!(x, g) #gradient evaluation to pre-allocated array
    g[1] = 8x[1] - 8.4x[1]^3 + 2x[1]^5 + x[2]
    g[2] = x[1] - 8x[2] + 16x[2]^3
    return nothing
end

function camel_fg!(x, g) #function call and gradient combined
    g[1] = 8x[1] - 8.4x[1]^3 + 2x[1]^5 + x[2]
    g[2] = x[1] - 8x[2] + 16x[2]^3
    return 4x[1]^2 - 2.1x[1]^4 + 1/3*x[1]^6 + x[1]*x[2] - 4x[2]^2 + 4x[2]^4
end

df = Optim.DifferentiableFunction(camel_f, camel_g!, camel_fg!)

result = optimize(df, [-5, -5], [5, 5]; show_trace=true)

The output is of type FminfinderOptimizationResults with following methods defined:

• Optim.minimizer: vector of points (each again a vector) where function reaches a minimum
• Optim.minimum: vector of function values at those points
• Optim.converged: whether a local search has converged at least once
• Optim.lower_bound, Optim.upper_bound, Optim.method, Optim.f_calls, Optim.g_calls : self explanatory

Additional options are:

• Ninit: initial number of sample points in the search space per iteration (default 20).
• Nmax: maximum number of sample points in the search space per iteration (default 100 as in the paper, but I usually find 250 more appropriate).
• enrich: multiplication of sample points N when more than half of sample points failed preselection criteria (default= 1.1).
• exhaustive: in (0,1). For small values p→0 the algorithm searches the area exhaustively, while for p→1 the algorithm terminates earlier, but perhaps prematurely (default value 0.5).
• max_algo_steps: maximum number of iterations, each with N points samples and local searches (default 10_000).
• show_trace: print iteration results (default = true)
• dist_unique: the results of a local search will be added to the minima list if its location differs less than this threshold from previously found minima. Increase when lots of returned minima correspond to the same physical point (default = sqrt(local_tol)).
• polish: boolean flag to indicate whether to perform an extra search at the very end for each minima found, to polish off the found minima with extra precision (default = true).
• distpolish: same as distmin for final polish phase (default sqrt(polish_tol) ).
• local_xtol: tolerance level used for the local searches, default eps(Type) (or sqrt(eps(T)^(2/3)) when polish=true). Similar poslish_ftol and polish_gtol (default eps(Type)^(2/3), or sqrt(eps(T)^(2/3)) when polish=true)
• polish_xtol: tolerance level of final polish searches (default eps(Type)^(2/3)).