[GA] InexactError when the target function returns Float and the inputs genes are Integers

[ndgnuh]

I met this issue while trying to use GA to solve the travelling salesman problem. The code to reproduce this issue is fairly simple, so I'll just paste it here:

using Evolutionary
using Random

struct TSPTargetFunction{T} <: Function
    weights::Matrix{T}
    function TSPTargetFunction(W::AbstractMatrix)
        new{eltype(W)}(W)
    end
end

function (f::TSPTargetFunction)(x)
    (
        f.weights[last(x), first(x)] +
        sum(f.weights[i, j] for (i, j) in zip(x, Iterators.drop(x, 1)))
    )
end

function solve_tsp(
    W::AbstractMatrix;
    population_size = 100,
    options_kwargs = Dict{Symbol,Any}(),
)
    f = TSPTargetFunction(W)
    options =
        Evolutionary.Options(iterations = 2000, successive_f_tol = 30, options_kwargs...)
    optimizer = GA(;
        populationSize = population_size,
        selection = rouletteinv,
        crossover = OX2, # not really, but not relevant
        mutation = x -> x[x], # not really, but not relevant
        mutationRate = 0.5,
        crossoverRate = 0.5,
        epsilon = population_size ÷ 5,
    )
    n = size(W, 1)
    initial_population = [randperm(n) for _ = 1:population_size]
    Evolutionary.optimize(f, initial_population, optimizer, options)
end

solve_tsp(rand(4, 4)) # Error

[wildart]

Yep, it's a bug. The initialization of the NonDifferentiable should be done better. Currently, the result of the objective function must have the same type as an element of the individual of the population. Otherwise, multiple functions has to be overridden to make it work. That is wrong.

• Ref: https://github.com/wildart/Evolutionary.jl/blob/master/src/api/types.jl#L117
• Solution: Use a result type of the objective function to initialize zero value

[ndgnuh]

Thanks, this workds for me (and shouldn't break the API)

function initialize_objective(f, x::AbstractArray{T}, S = typeof(f(x))) where T
    NonDifferentiable{S,typeof(x)}(f, zero(S), zeros(T, length(x)), [0,])
end

initialize_objective(f, first(individual))