Problem with optimization with Optim.jl

[STaherpour]

Hi, I'm trying to use Optim.jl in Optimization of a State-to-State Transfer in a Two-Level-System.
But when I run this part of the code:

opt_result_OptimLBFGS = @optimize_or_load(
    datadir("TLS", "opt_result_OptimLBFGS.jld2"),
    problem,
    method = :grape,
    info_hook = chain_infohooks(
        GRAPELinesearchAnalysis.plot_linesearch(datadir("TLS", "Linesearch", "OptimLBFGS")),
        QuantumControl.GRAPE.print_table,
    ),
    optimizer = Optim.LBFGS(;
        alphaguess=LineSearches.InitialStatic(alpha=0.2),
        linesearch=LineSearches.HagerZhang(alphamax=2.0)
    )
);

I receive this error:

┌ Warning: Using the standard Julia project.
└ @ DrWatson /root/.julia/packages/DrWatson/I8SbQ/src/project_setup.jl:30
┌ Warning: Using the standard Julia project.
└ @ DrWatson /root/.julia/packages/DrWatson/I8SbQ/src/project_setup.jl:30
┌ Warning: Using the standard Julia project.
└ @ DrWatson /root/.julia/packages/DrWatson/I8SbQ/src/project_setup.jl:30
 iter.        J_T     |∇J_T|       ΔJ_T   FG(F)    secs
     0   9.51e-01   5.40e-02        n/a    1(0)     6.8
     1   9.51e-01   5.40e-02   0.00e+00    3(0)     2.0
     2   9.51e-01   5.40e-02   0.00e+00    3(0)     4.8
┌ **Warning: Optimization failed to converge**
└ @ GRAPE /root/.julia/packages/GRAPE/WVcll/src/backend_optim.jl:87
┌ Warning: Using the standard Julia project.
└ @ DrWatson /root/.julia/packages/DrWatson/I8SbQ/src/project_setup.jl:30
┌ Warning: Using the standard Julia project.
└ @ DrWatson /root/.julia/packages/DrWatson/I8SbQ/src/project_setup.jl:30
┌ Warning: The directory ('/root/.julia/environments/v1.8') is not a Git repository, returning `nothing` instead of the commit ID.
└ @ DrWatson /root/.julia/packages/DrWatson/I8SbQ/src/saving_tools.jl:52

Would you please help me in this regard? @goerz

[goerz]

I'll have a look when I have a chance (which might take some time). In the meantime, I recommend sticking to the default L-BFGS-B optimizer.

[STaherpour]

Thank you so much.
Actually I want to learn working with Optim.jl and my main issue is working with Optim.GradientDescent.

When I use

opt_result_OptimGD = QuantumControlBase.optimize(
    
    problem,
    method = :grape,
    info_hook = chain_infohooks(
        QuantumControl.GRAPE.print_table,
        store_pulses
    ),
    optimizer = Optim.GradientDescent(;
        #alphaguess=LineSearches.InitialStatic(alpha=0.2),
        alphaguess = LineSearches.InitialPrevious(alphamin=0.2),
        linesearch = LineSearches.HagerZhang(alphamax=2.0)
    )
);

I face this error: type GradientDescentState has no field g_previous, and I don't know why it happens. The g_previous is in the Optim.LBFGS not in the Optim.GradientDescent.

[goerz]

Yes, LBFGS was the only supported method. I'll make gradient descent work in the next release.

[STaherpour]

I just find out that in GRAPELinesearchAnalysis.jl there is _get_linesearch_data function which needs optimizer_state.g_previous. But when I use Optim.jl for optimization, and select Optim.GradientDescent, it doesn't have any g_previous. instead it has x_previous.
I think it makes problem. But unfortunately I don't know how to manage it.

function _get_linesearch_data(
    wrk,
    optimization_state::Optim.OptimizationState,
    optimizer_state::Optim.AbstractOptimizerState,
)
    # TODO: use wrk only (and compare with current implementation)
    α = optimizer_state.alpha # the stepwidth chosen by the linesearch
    lbfgs_direction = optimizer_state.s # the quasi-Newton update direction
    gradient_direction = -1 * optimizer_state.g_previous
    pulsevals_guess = optimizer_state.x_previous
    pulsevals_opt = optimizer_state.x

    return α, lbfgs_direction, gradient_direction, pulsevals_guess, pulsevals_opt

function GradientDescent(; alphaguess = LineSearches.InitialPrevious(), # TODO: Investigate good defaults.
                           linesearch = LineSearches.HagerZhang(),      # TODO: Investigate good defaults
                           P = nothing,
                           precondprep = (P, x) -> nothing,
                           manifold::Manifold=Flat())
    GradientDescent(_alphaguess(alphaguess), linesearch, P, precondprep, manifold)
end

mutable struct GradientDescentState{Tx, T} <: AbstractOptimizerState
    x::Tx
    x_previous::Tx
    f_x_previous::T
    s::Tx
    @add_linesearch_fields()
end

[goerz]

Likewise, there was no support for GradientDescent, but it should work in the next release

[goerz]

What you're seeing with Optimization failed to converge is a bug; see JuliaQuantumControl/GRAPE.jl#38. This will be fixed in the next release.

Since there's a major release of all packages in the pipeline, it might take a couple of weeks before this becomes available. In the meantime, you could apply the patch to your local installation.

[STaherpour]

Thanks a lot