JuMP model optimisation

PortfolioOptimisers.jl is based on JuMP, as such it tries to be as flexible as possible.

These types and functions let us define solver and solution interfaces.

PortfolioOptimisers.Solver — Type

struct Solver{__T_name, __T_solver, __T_settings, __T_check_sol, __T_add_bridges} <: AbstractEstimator

Configures one solver backend, its attributes, and the statuses its solutions must reach.

Every optimiser takes one Solver or a vector of them. optimise_JuMP_model! tries them in order and stops at the first that returns a solution check_sol accepts, so a vector is a fallback chain.

Fields

name: Symbol or string identifier. It is also the key under which optimise_JuMP_model! files this solver's failure in JuMPResult's trials, so two solvers that share a name share one entry and the later failure overwrites the earlier one. Give each solver of a vector its own name, because the default "" is shared by all of them.

solver: The optimizer_factory in set_optimizer.

settings: Optional solver-specific settings used in set_attribute.

check_sol: Named tuple of keyword arguments splatted into assert_is_solved_and_feasible after each solve. It decides which solver statuses count as a solved model. The default (;) accepts JuMP's own defaults, allow_local = true and allow_almost = false: the termination status must be OPTIMAL or LOCALLY_SOLVED, and the primal status must be FEASIBLE_POINT. The strictness is deliberate — a solution the solver itself flags as approximate is rejected rather than silently accepted, so a solver stage that fails this check falls through to the next solver in the vector. The common relaxation is check_sol = (; allow_local = true, allow_almost = true), which also accepts ALMOST_OPTIMAL, ALMOST_LOCALLY_SOLVED and NEARLY_FEASIBLE_POINT; it is what the examples, the user guide and the test suite pass, because a first-order solver reaching its tolerance on a well-posed portfolio problem is usually good enough. Go the other way with allow_local = false to reject LOCALLY_SOLVED and demand a certified global optimum, and add dual = true to also require a feasible dual point.

add_bridges: The add_bridges keyword argument in set_optimizer.

Constructors

Solver(;    name::Sym_Str = "",    solver::Any,    settings::Option{<:SlvSettings} = nothing,    check_sol::NamedTuple = (;),    add_bridges::Bool = true) -> Solver

Keywords correspond to the struct's fields.

Validation

If not nothing, !isempty(settings).

Examples

The default check_sol is strict — it accepts only a solution the solver reports as OPTIMAL or LOCALLY_SOLVED at a FEASIBLE_POINT.

julia> Solver(; solver = nothing)Solver         name ┼ String: ""       solver ┼ nothing     settings ┼ nothing    check_sol ┼ @NamedTuple{}: NamedTuple()  add_bridges ┴ Bool: true

To also accept an approximate solution, which is the common case and what the examples, user guide and tests use:

julia> Solver(; solver = nothing, check_sol = (; allow_local = true, allow_almost = true))Solver         name ┼ String: ""       solver ┼ nothing     settings ┼ nothing    check_sol ┼ @NamedTuple{allow_local::Bool, allow_almost::Bool}: (allow_local = true, allow_almost = true)  add_bridges ┴ Bool: true

Related

source

PortfolioOptimisers.JuMPResult — Type

struct JuMPResult{__T_trials, __T_success} <: AbstractJuMPResult

Records which solvers failed, at which stage, and whether any of them succeeded.

When success is false the constructor emits a warning built by failed_solve_msg: one bounded line per failed solver stage (name, stage, first line of the error). The full per-solver exceptions and settings stay available on trials and are never dumped into the log.

trials records failures only. A solver that succeeds leaves no entry, so an empty trials with success = true means the first solver answered, and it is not a record of the solve. Each entry is keyed by the solver's name and holds a dictionary from the failed stage — :set_optimizer, :optimize! or :assert_is_solved_and_feasible — to the exception. Two solvers that share a name share one entry, and the default name is "" for all of them, so a vector of unnamed solvers keeps only its last failure.