optimizers_interface.jl

using Optim, BlackBoxOptim, NLopt, PyCall

import BlackBoxOptimizationBenchmarking: minimizer, minimum, optimize 
import Base.string

box(D) = fill((-5.5, 5.5), D)
pinit(D) = 10*rand(D).-5

# Define optimize, minimum and minimizer for each optimizer

## NLopt

    mutable struct NLoptOptimMethod 
        s::Symbol
    end
    string(opt::NLoptOptimMethod) = string("NLopt.", opt.s)

    function optimize(NLmeth::NLoptOptimMethod, f, D, run_length)  
        opt = Opt(NLmeth.s, D)
        min_objective!(opt, (p, g) -> f(p) ) #NLopt expect gradient
        maxeval!(opt, run_length)
        xtol_abs!(opt, 1e-12)
        xtol_rel!(opt, 1e-12)
        ftol_abs!(opt, 1e-12)
        ftol_rel!(opt, 1e-12)
        lower_bounds!(opt, -5.5*ones(D))
        upper_bounds!(opt, +5.5*ones(D))
        minf, minx, ret = NLopt.optimize(opt, pinit(D))
        return NLmeth, minx, minf
    end
    minimum(mfit::Tuple{NLoptOptimMethod, Vector{Float64}, Float64}) = mfit[3]
    minimizer(mfit::Tuple{NLoptOptimMethod, Vector{Float64}, Float64}) = mfit[2]

## chain

    mutable struct Chain{T, K}
        first::T
        second::K
        p::Float64
    end
    
    function optimize(m::Chain, f, D, run_length) 
        rl1 = round(Int, m.p*run_length)
        rl2 = run_length - rl1
        
        mfit = optimize(m.first, f, D, run_length)
        xinit = minimizer(mfit)
        mfit = optimize(m.second, f, D, run_length, xinit) 
    end
    
    string(m::Chain) = string(string(m.first), " → ", string(m.second))
    

## python cma

#try

    @pyimport cma
    
    struct PyCMA
    end

    function optimize(m::PyCMA, f, D, run_length)
        es = cma.CMAEvolutionStrategy(pinit(D), 3, Dict("verb_log"=>0, "verb_disp"=>0, "maxfevals"=>run_length))
        mfit = es.optimize(f).result
        (m, mfit[1], mfit[2])
    end

    string(m::PyCMA) = "PyCMA"
    minimum(mfit::Tuple{PyCMA, Vector{Float64}, Float64}) = mfit[3]
    minimizer(mfit::Tuple{PyCMA, Vector{Float64}, Float64}) = mfit[2]
  
## scipy

    @pyimport scipy.optimize as scipy_opt

    struct PyMinimize
        method::String
    end

    optimize(m::PyMinimize, f, D, run_length) = (m, scipy_opt.minimize(
        f, pinit(D), method=m.method,
        options = Dict(
            "maxfev"=>run_length, "xatol"=>1e-8, "fatol"=>1e-8,
            "maxiter"=>run_length, "gtol"=>1e-12,
        )
    ))
    minimum(mfit::Tuple{PyMinimize, Dict{Any, Any}}) = mfit[2]["fun"]
    minimizer(mfit::Tuple{PyMinimize, Dict{Any, Any}}) = mfit[2]["x"]
    
    string(m::PyMinimize) = string("Py.", m.method)
    
#end
    
## Optim

    optimize(opt::Optim.AbstractOptimizer, f, D, run_length) =
        Optim.optimize(f, pinit(D), opt, Optim.Options(f_calls_limit=run_length, g_tol=1e-120, iterations=run_length))
                
    optimize(opt::Optim.AbstractOptimizer, f, D, run_length, xinit) =
        Optim.optimize(f, xinit, opt, Optim.Options(f_calls_limit=run_length, g_tol=1e-120, iterations=run_length))
        
    optimize(opt::Optim.SAMIN, f, D, run_length) =
        Optim.optimize(
            f, fill(-5.5, D), fill(5.5, D), pinit(D), opt, 
            Optim.Options(f_calls_limit=run_length, g_tol=1e-120, iterations=run_length, show_trace=false)
        )

    string(opt::Optim.AbstractOptimizer) = string(typeof(opt).name)

    # Optim with restart
    try 
        mutable struct OptimRestart{T}
            opt::T
        end
    catch
    end

    function optimize(opt::OptimRestart, f, D, run_length) 
        fits = [optimize(opt.opt, f, D, round(Int, run_length/20)) for i=1:20]
        mins = [minimum(fit) for fit in fits]
        fits[argmin(mins)]
    end

    string(opt::OptimRestart) = string("Restart-", string(opt.opt))

## BlackBoxOptim

    mutable struct BlackBoxOptimMethod 
        s::Symbol
    end
    string(opt::BlackBoxOptimMethod) = string("BBO.", opt.s)
    
    optimize(method::BlackBoxOptimMethod, f, D, run_length) =
        bboptimize(f; SearchRange=box(D), NumDimensions=D, Method=method.s, MaxFuncEvals=run_length, TraceMode=:silent)

    minimum(mfit::BlackBoxOptim.OptimizationResults) = best_fitness(mfit)
    minimizer(mfit::BlackBoxOptim.OptimizationResults) = best_candidate(mfit)