Skip to content

add option to change subsolver in AL #310

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
MaxenceGollier wants to merge 1 commit into
base: master
Choose a base branch
from
Open

add option to change subsolver in AL #310

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 8 additions & 6 deletions src/AL_alg.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -81,7 +81,7 @@ where y is an estimate of the Lagrange multiplier vector for the constraints lco

For advanced usage, first define a solver "ALSolver" to preallocate the memory used in the algorithm, and then call `solve!`:

solver = ALSolver(reg_nlp)
solver = ALSolver(reg_nlp; subsolver = R2Solver)
solve!(solver, reg_nlp)

stats = GenericExecutionStats(reg_nlp.model)
Expand All @@ -108,7 +108,7 @@ If adopted, the Hessian is accessed as an abstract operator and need not be the
- `max_iter::Int = 10000`: maximum number of iterations;
- `max_time::Float64 = 30.0`: maximum time limit in seconds;
- `max_eval::Int = -1`: maximum number of evaluation of the objective function (negative number means unlimited);
- `subsolver::AbstractOptimizationSolver = has_bounds(nlp) ? TR : R2`: the procedure used to compute a step (e.g. `PG`, `R2`, `TR` or `TRDH`);
- `subsolver::AbstractOptimizationSolver = R2Solver`: the procedure used to compute a step (e.g. `R2Solver`, `R2NSolver`, `R2DHSolver`, `TRSolver` or `TRDHSolver`);
- `subsolver_logger::AbstractLogger`: a logger to pass to the subproblem solver;
Comment on lines 110 to 112
Copy link

Copilot AI Mar 9, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The docstring lists subsolver::AbstractOptimizationSolver = R2Solver, but subsolver is used as a constructor/callable (e.g. subsolver(reg_nlp; ...)) rather than an AbstractOptimizationSolver instance. To avoid confusing users (and to match how other solvers document subsolver), update the doc to reflect that this expects a solver type/constructor (e.g. Type{<:AbstractOptimizationSolver} or a callable that returns a solver) and clarify the default accordingly.

Copilot uses AI. Check for mistakes.
- `init_penalty::T = T(10)`: initial penalty parameter;
- `factor_penalty_up::T = T(2)`: multiplicative factor to increase the penalty parameter;
Expand Down Expand Up @@ -148,7 +148,7 @@ mutable struct ALSolver{T, V, M, Pb, ST} <: AbstractOptimizationSolver
sub_stats::GenericExecutionStats{T, V, V, T}
end

function ALSolver(reg_nlp::AbstractRegularizedNLPModel{T, V}; kwargs...) where {T, V}
function ALSolver(reg_nlp::AbstractRegularizedNLPModel{T, V}; subsolver = R2Solver, kwargs...) where {T, V}
nlp = reg_nlp.model
nvar, ncon = nlp.meta.nvar, nlp.meta.ncon
x = V(undef, nvar)
Expand All @@ -157,7 +157,7 @@ function ALSolver(reg_nlp::AbstractRegularizedNLPModel{T, V}; kwargs...) where {
has_bnds = has_bounds(nlp)
sub_model = AugLagModel(nlp, V(undef, ncon), T(0), x, T(0), cx)
sub_problem = RegularizedNLPModel(sub_model, reg_nlp.h, reg_nlp.selected)
sub_solver = R2Solver(reg_nlp; kwargs...)
sub_solver = subsolver(reg_nlp; kwargs...)
sub_stats = RegularizedExecutionStats(sub_problem)
Comment on lines 157 to 161
Copy link

Copilot AI Mar 9, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ALSolver builds sub_problem as an augmented-Lagrangian RegularizedNLPModel, but the sub_solver is constructed using reg_nlp (the original problem). For subsolvers that precompute model-dependent operators/state (e.g. R2NSolver builds a Hessian operator in its constructor), this means the subsolver is initialized for the wrong model and then used to solve sub_problem, which can lead to incorrect steps or failures. Construct the subsolver from sub_problem instead (and forward any subsolver-specific kwargs there) so the solver is consistent with the subproblem it will be asked to solve.

Copilot uses AI. Check for mistakes.
Copy link
Member

@dpo dpo Mar 11, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copilot is right, here, no?

M = typeof(nlp)
ST = typeof(sub_solver)
Expand All @@ -182,8 +182,10 @@ end
"AL(::Val{:equ}, ...) should only be called for equality-constrained problems with bounded variables. Use AL(...)",
)
end
solver = ALSolver(reg_nlp)
solve!(solver, reg_nlp; kwargs...)
kwargs_dict = Dict(kwargs...)
subsolver = pop!(kwargs_dict, :subsolver, R2Solver)
solver = ALSolver(reg_nlp, subsolver = subsolver)
solve!(solver, reg_nlp; kwargs_dict...)
end

function SolverCore.solve!(
Expand Down
9 changes: 8 additions & 1 deletion test/test_AL.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,14 @@ problem_list = [:hs8]
for problem in problem_list
nlp = eval(problem)(backend = :optimized)
for h in (NormL1(1.0),)
stats = AL(nlp, h, atol = 1e-3, verbose = 1)
stats = AL(nlp, h, subsolver = R2Solver, atol = 1e-3, verbose = 1, subsolver_max_iter = 1000)
@test stats.status == :first_order
@test stats.primal_feas <= 1e-2
@test stats.dual_feas <= 1e-2
@test length(stats.solution) == nlp.meta.nvar
@test typeof(stats.solution) == typeof(nlp.meta.x0)

stats = AL(LBFGSModel(nlp), h, subsolver = R2NSolver, atol = 1e-3, verbose = 1, subsolver_max_iter = 1000)
@test stats.status == :first_order
@test stats.primal_feas <= 1e-2
@test stats.dual_feas <= 1e-2
Expand Down
Loading