Uno v2.0.0

[cvanaret]

Uno v2.0.0 implements the following changes:

Features

Abstract framework

• The underlying abstract framework was refined and is now more powerful. Uno now implements 8 ingredients: constraint relaxation strategy, inequality handling method, Lagrange-Newton subproblem, Hessian model, subproblem solver, regularization strategy, globalization strategy and globalization mechanism (Ingredient refactoring #122, [WIP] Introduction of Subproblem class #183)

• Replaced LP/QP subproblem with generic inequality-constrained method. The subproblem solver (LP or QP) is automatically allocated, depending whether the subproblem has some curvature (Automatically allocate LP/QP solver for inequality-constrained method #177)
• New constraint relaxation strategy UnconstrainedStrategy automatically set for unconstrained problems (Special treatment for unconstrained problems #161)
• The HessianModel is now a model of the original Hessian (Moved the Hessian model between the Model and the reformulations #152)
• The combination name is generated automatically. Example: TR Fletcher-filter restoration inequality-constrained method with exact Hessian and no regularization (Generation of the combination name #179)

User friendliness

• User callbacks can be provided to the Uno::solve() call (Add user callbacks #106, Suggestion for implementing user callbacks #108)
• The solution status was split into iterate status and optimization status (Separated iterate status and optimization status #110)

Interfaces

• Added an interface to the symmetric indefinite linear solver MA27 (MA27 interface #115, Review of the MA27 interface #116, MA27: upper bound on factorization attempts #119)

Continuous integration:

• Build Uno with MSVC on Windows (Windows MSVC support? #144, Build Uno with MSVC on windows #146)
• Cross-compile Uno with the indefinite QP solver BQPD. BQPD is now available as pre-compiled binaries (Cross-compile Uno with BQPD #168, Test Uno with BQPD for all platforms #173)
• Cross-compile Uno with HiGHS v1.11 (Test HiGHS v1.11 #170)
• Cross-compile Uno with MUMPS 5.8.0 ([CI] Compile Uno with MUMPS #190)

Improvements:

• Improved the scaling in the MUMPS interface (Reduce the number of symbolic analyses #104, MUMPSSolver: perform the symbolic analysis without matrix entries #120)
• Uno::solve() now returns a Result object, even when the optimization fails (Uno::solve() should return the result #105, Return the Result from Uno::solve() #111)
• Compute symbolic analyses only when the Hessian sparsity changes (Reduce the number of symbolic analyses #104, Fixed memory allocation in IPM + removed useless symbolic analyses #129)
• Improve the memory allocations (Improve memory allocation #155, Fix the number of nonzeros in the regularized Hessian/augmented system #172)
• The Hessian is not saved in the BQPD workspace any more (BQPDSolver: moved the gdotx function from Fortran to C++ #135, BQPD: use the existing sparse Hessian instead of copying it in the workspace #191)
• The Uno class is now instantiated with an Options object and the number of constraints. The actual memory allocation and possible choices of strategies are performed when Uno::solve() is called (Improve the construction of the Uno object #36)
• Improved the switch to the feasibility phase in FeasibilityRestoration (FeasibilityRestoration: move the switch from phase 1 to phase 2 to the end #180)
• Automatically deduce the indices of variables and constraints that should be regularized, respectively for primal and dual regularization (Compute the (primal and dual) regularization indices automatically #187)
• Split the barrier method into the method itself and the nonlinear barrier reformulation (Split primal-dual interior-point reformulation and method #126, Improve genericity of barrier reformulation #188)
• Improve the design of symmetric matrices: SparseSymmetricMatrix<SparseStorage> now inherits SymmetricMatrix. This is in preparation of the Python bindings ([WIP] Improve the matrix hierarchy #194)

Bugfixes:

• Fixed the computation of the complementarity residual in l1RelaxedProblem (Fixed complementarity in l1RelaxedProblem #149)
• Fixed warmstart information in TrustRegionStrategy (Fixed warmstart information in TrustRegionStrategy #174)
• Fixed memory allocation in Preprocessing::compute_least_square_multipliers() (Allocate own matrix and RHS in Preprocessing::compute_least_square_multipliers() #176)
• Fixed the multipliers (whose length may be larger than expected) in AMPLModel (Bugfix in AMPLModel: size of multipliers vector may be larger than expected #184)
• Fixed cross-compilation on Apple Darwin (Fixed missing include (Apple Darwin) #203, Fixed missing include (Apple Darwin) #204, Use the old linker on Mac #205, Adapt CMAKE_Fortran_IMPLICIT_LINK_LIBRARIES for Mac #206, Use ld instead of lld for Mac-Intel #208)

Acknowledgments

Big thanks go to Alexis Montoison (@amontoison) for his help during the last steps of the release.