API reference

TRResult

An alias of JSO's GenericExecutionStats. Useful fields: status, solution, objective, dual_feas, iter, elapsed_time. Evaluation counts live on the model — neval_obj(nlp), neval_grad(nlp), neval_hprod(nlp).

AbstractTRSolver <: AbstractOptimizationSolver

Supertype of the three solvers. Which one applies is decided by the problem class, not by a keyword:

tr_solve dispatches for you.

tr_solve(nlp; rule, model, subsolver, params, x_ref, true_curvature,
         hessian_norm, kwargs...) -> TRResult

The declared keywords are rule, model, subsolver, params, x_ref, true_curvature and hessian_norm. Everything else, trace and callback above all, is forwarded through kwargs... to SolverCore.solve!, so trace = true is the supported spelling rather than a keyword of this function.

Solve nlp, choosing the solver from the problem class:

The sampling rule is not an argument here: it belongs to the oracle, because it decides what the oracle returns.

using TrustRegionRadius, ADNLPModels

# deterministic
nlp = ADNLPModel(x -> (1 - x[1])^2 + 100(x[2] - x[1]^2)^2, [-1.2, 1.0])
tr_solve(nlp; rule = RGrad(), model = SR1Model(mem = 5))

# finite sum: the cap is M, not a keyword
prob = LogisticRegression(K = 5, M = 2_000)
tr_solve(FiniteSumNLP(prob, RadiusProportional()); model = BHHHModel(ridge = 1e-8))

# expectation: the cap is the budget, because there is no M
exp_prob = PerturbedExpectation(base; σg = 1.0)
tr_solve(ExpectationNLP(exp_prob, RadiusProportional(); budget = 50_000))

model_grad_evals(solver) -> Int

Gradient evaluations made from inside the model, which the algorithm did not request. Non-zero only for SPDTarget, which calls grad inside dense_hessian and phi_target. Subtract before reporting neval_grad as a cost:

algorithm_grads = neval_grad(nlp) - model_grad_evals(solver)

The fallback returns 0 because no other model evaluates the gradient itself.