adapt to new SurrogatesBase and improve performance of add_points

[jbrea]

No description provided.

[samuelbelko]

Thanks for your contributions, what version of SurrogatesBase are you referring to? Is it intentional to omit "at point" versions of mean, var, mean_and_var, add_point! ?

[samuelbelko]

This is cool, thanks!

[samuelbelko]

👍

[samuelbelko]

In the posterior, we should use the whole history g.xs, g.ys, not just new_xs, new_ys.
The problem why we have to copy the history is that once we create a posterior, we cannot change the data. I think we cannot avoid copying the history each time we add new data, since we update the posterior each time.

[samuelbelko]

Now, I am a bit confused about the error.

The following does not throw an error:

function kernel_creator(hyperparameters)
    return with_lengthscale(KernelFunctions.Matern52Kernel(), hyperparameters.lengthscale)
end
@testset "test" begin
    xs= [1.0, 2.0, 3.0]
    ys = [1., 2., 3.]
    l = GPSurrogate(xs, ys, kernel_creator = kernel_creator,
    hyperparameters = (; lengthscale = 1.0, noise_var = 0.1))
    push!(l.xs, 10.)
    push!(xs, 20.)
    println(l.xs)
    @show l.gp_posterior
end

But when I don't pass the keyword arguments,

@testset "test" begin
    xs= [1.0, 2.0, 3.0]
    ys = [1., 2., 3.]
    l = GPSurrogate(xs, ys)
    push!(l.xs, 10.)
    push!(xs, 20.)
    println(l.xs)
    @show l.gp_posterior
end

it throws cannot resize array with shared data.

[jbrea]

In the posterior, we should use the whole history g.xs, g.ys

I think the whole history is taken into account. Note that I'm updating the posterior, instead of recomputing the posterior from scratch, but this should be tested.

Copying the history and recomputing the posterior is insanely slow. If a fast alternative is not possible with AbstractGPs, we modify AbstractGPs or we use GaussianProcesses.jl, where I use the fast ElasticPDMats.

[samuelbelko]

Alright thanks! I found out in the AbstracGPs docs that it is possible to do squential conditioning. (Available here)

[samuelbelko]

Would it be better to use copied input in the surrogate? It is probably safer to do so, avoiding mutations from outside. 👍

[jbrea]

This is just for the construction. We can also have copying optional with a keyword argument.

[jbrea]

what version of SurrogatesBase are you referring to?

samuelbelko/SurrogatesBase.jl#1

[samuelbelko]

I merged all proposed changes while adapting to recent changes of SurrogatesBase. Copying of the arguments is not yet optional though.