Surrogate model

The surrogate model which is used for the optimization can be created manually with

SurrogateModelOptim.surrogate_model — Function

surrogate_model(plan::AbstractArray{T,2}, samples::AbstractArray{T,2}; options=Options()) where T

Returns a surrogate model function based on an optimized Radial Basis Function interpolant. Depending on the supplied options; the kernel, kernel width and scaling of input data is optimized.

...

Arguments

plan::AbstractArray{T,2}: sample locations where each column corresponds to the location of one point. size(plan) = (num_dimensions,num_samples).
samples::AbstractArray{T,2}: function value at each sample location. each column contains one value from the corresponding plan location. size(samples) = (1,num_samples).
options=Options(): all options available to customize the surrogate optimization.

...

source

This enables the freedom to choose how it is optimized and used. When called, the function value from each surrogate in the ensemble is returned.

Example

julia> rosenbrock_2D(x) = (1.0 - x[1])^2 + 100.0 * (x[2] - x[1]^2)^2
julia> search_range=[(-5.0,5.0),(-5.0,5.0)]

# Start from 5 Latin Hypercube Samples
julia> num_samples=5
julia> sampling_plan_opt_gens=10_000
julia> plan = scaled_LHC_sampling_plan(search_range,num_samples,sampling_plan_opt_gens)

# Evaluate the function 
julia> samples = mapslices(rosenbrock_2D,plan,dims=1)

# Create the optimized surrogate model (optres contains the optimization results for the surrogate)
julia> opt=SurrogateModelOptim.Options()
julia> sm_interpolant, optres = surrogate_model(plan, samples;options=opt)

The surrogate model is by default created using only Gaussian kernels. This can be changed by supplying several kernels, e.g. kerns = [ScatteredInterpolation.Gaussian, ScatteredInterpolation.InverseQuadratic, ScatteredInterpolation.InverseMultiquadratic]. For a limited number of rbf_opt_gens the surrogate model performance tends to be better using only one kernel to limit the search space when optimizing the surrogate hyperparameters.

Turning on verbosity with trace=:verbose can help with judging of the hyperparameter optimisation.