Pkg.clone("https://github.com/gsoleilhac/NSGAIII.jl")
popsize = 200
nbGenerations = 100
#Define the number of division along each objective to generate the reference directions.
H = 5
#Alternatively, you can directly pass the reference directions as a Vector{Vector{Float64}} :
#With two objectives, H = 2 is equivalent to
H = [[1., 0.], [0.5, 0.5], [0., 1.]]
#define how to generate a random genotype :
init_function = () -> randperm(N) #e.g : a permutation coding
#define how to evaluate a genotype :
z(x) = z1(x), z2(x) ... # Must return a Tuple
nsga(popsize, nbGenerations, init_function, z, H)
#A constraint violation function can be passed with the keyword fCV
#It should return 0. if the solution is feasible and a value > 0 otherwise.
#Mutation probability can be changed with the keyword pMut (default is 5%)
nsga(popsize, nbGen, init_fun, z, H, fCV = CV, pMut = 0.1)
By default, a PMX Crossover and a random swap mutation will be used if the genotype is a Vector{Int}
and a two-point crossover and random flips will be used for Vector{Bool} / BitArrays.
Other crossover / mutations operators can be passed with the keywords fmut and fcross
nsga(popsize, nbGen, init, z , fmut = ..., fcross = ...)
Starting solutions are defined with the keyword seed
e.g :
x1, x2, x3 = greedy(...)
nsga(popsize, nbGen, init, z, seed = [x1, x2, x3])
A plot function can be passed with the keyword fplot
e.g. with PyPlot:
function plot_pop(P)
clf()
p = plot(map(x -> x.y[1], P), map(x -> x.y[2], P), "bo", markersize=1)
sleep(0.2)
end
nsga(popsize, nbGen, init, z, fplot = plot_pop)
RealCoding(eps, lb, ub) and decode(x, d::RealCoding) can be used to easily represent Real values with eps decimal places.
encode(x, d) can be used to provide starting solutions
d = RealCoding(6, [-3, -3], [6, 6]) #Codes two reals with 6 decimal places with lower bound -3 and upper bound 6
z1(x1, x2) = -(3(1-x1)^2 * exp(-x1^2 - (x2+1)^2) - ...)
z2(x1, x2) = -(3(1+x2)^2 * exp(-x2^2 - (1-x1)^2) - ...)
z(x) = begin
x1, x2 = decode(x, d)
z1(x1, x2), z2(x1, x2)
end
x1, x2 = (-3., 6), (-$\pi$, 0)
seed = encode.([x1, x2], d)
nsga(100, 50, ()->bitrand(d.nbbitstotal), z, seed=seed)
This package supports models declared with JuMP / vOptGeneric, with the restriction that all variables must be bounded, and that they're either continuous or binary (integer might get added later)
See examples/Mavrota_MILP.jl