NEWS.md

Version 0.8.0 (02. February 2021)

This is a fairly big release as we wrap the ADMM algorithm into a safeguarded Anderson Acceleration method. The accelerators are provided by COSMOAccelerators.jl. This should improve the convergence to higher accuracy solution.

We change the default settings to reflect this. The default accuracies eps_abs and eps_rel are set to 1e-5, we now check convergence every 25 iterations, and increased max_iter to 5000. Acceleration can be switched off by using: COSMO.Settings(accelerator = EmptyAccelerator). When the algorithm is safeguarded (safeguard = true), extra ADMM-iteration called safeguarding_iter might occur which are also reported to the user. The returned number of iterations in the result object always refers to the total number of ADMM-iterations, which includes the safeguarding_iters. Refer to the docs for more configuration options.

• fb51a41 Add Compat entry for Reexport
• 2624a50 Bump version: 0.7.9 → 0.8.0
• 58cc506 Update examples/ to newest JuMP API
• ab67ddc Fix bug in ResulType constructor
• 5ac07e9 Merge branch 'mg/accelerated_algorithm'

Version 0.7.9 (02. February 2021)

This is intended as the last release containing all changes before the accelerated version of the solver is merged.

• a8f2724 Merge pull request #129 from oxfordcontrol/mg/fix_slow_moi_merging
• 527c5d2 Increase max_iter and decrease eps_prim_inf
• b4ea038 Change Int64 --> Int in clique graph test
• 0744299 [#124] Covering more cases in CI
• 69d8e4f Export CGIndirectKKTSolver and MINRESIndirectKKTSolver
• 1ef8ca7 Update Documenter version
• 7e1e0b0 Exclude solve_farkas_interval_upper from MOI unittests
• 19f43ec 👷 Switch CI to Github actions

Version 0.7.8 (21. November 2020)

• f0070a1 Bump DataStructures.jl to support v0.18 (#119)
• fd5b2b5 Add doc for model updates
• 316e328 Fix bracket bug in cost computation
• ef731a4 Rename update!(strategy,...) to update_strategy!(strategy,...)
• cbe6cc3 Add update!-model to docs
• 04283c2 Handle the model state more precisely
• 5db323f Fix slow Pardiso KKT solver

Version 0.7.7 (28. October 2020)

Fixes bugs in the interface functions for the Python interface.

Version 0.7.6 (25. October 2020)

This version adds some interface functions specifically for the Python interface: cosmo-python.

Version 0.7.5 (17. September 2020)

• This version implements the reduced clique graph merge strategy which fixes some edge cases that occurred in the previous merge strategy. It also makes this COSMO version compatible with the latest MathOptInterface release v0.9.16

• 95dbdf9 <#116> Make COSMO compatible with MOI v0.9.16

• 8d869f5 Fix wrong bracket in Box support function check

• 91bb113 Free clique graph related memory after the merging

• 9729d3c Clean-up reduced clique graph code

• 09eb9e5 Make ParCh strategy work with sets

• 381db30 Implement reduced clique graph merge

Version 0.7.4 (3. August 2020)

• This version mostly contains minor changes and small bug fixes as well as some updates to the documentation.

• 6626c6d [#114] Restrict dependencies less

• 2dcf4a8 Print correct number of sets in verbose output

• 58ef719 Untrack Manifest.toml

• 9cdd856 Fix sign error in logistic regression example

• 5ea43c7 Update Documenter to latest version

• ce6a10a Fix navigation bar error

• 732f76c Move Logistic regression to Documentation

• 14e3182 Change logo in docs

Version 0.7.3 (15. June 2020)

• We significantly improved the solver performance in this release. This is mainly achieved by direct assembly of the KKT matrix in CSC format, faster permutation and getting rid of some function overhead related to the SplitVector type. Moreover, we now made the solver precision type-agnostic, which means you can use any AbstractFloat type, e.g. BigFloat for your problem data (some limitations apply, see docs) and COSMO will solve the problem with that type.

• 0b0f3b3 Always time iter_start

• 6ff52a8 Make chordal decomposition code type agnostic

• 3692d49 Make the MOI-wrapper precision-type-agnostic

• ae43c04 Add more strongly typed arguments

• 4801beb Make solver work for ArbitraryFloat data

• 8237358 🎨 Print result status in color

• 8277eb0 Add minor performance improvements

• 4a79217 Add fast lower triangle KKT assembly function

• 906b9a0 Removed asserts. Direct access to SplitVector data. Removed CSC matrix casts. Removed some zipped calls.

• a6aa495 isolate triu KKT build for speed

• 921e927 minor loop rewrites

• ae21123 fast triu csc assembly

• 0184138 faster scalings and sqrt

• 2a1d43a isolated Box scaling function

• b3c15ed faster norms and l/r scaling

• 55f4511 faster qdldl

Version 0.7.2 (22. March 2020)

• We now handle the optional dependencies of external linear solver packages Pardiso and IterativeSolvers by using the Requires.jl package. Furthermore, we improve convergence of problems with box constraints in many cases. We also allow an automatic adaptive rho interval option, like in OSQP, which can be activated by setting adaptive_rho_interval = 0 (automatic).

• The main commits in this new version are:

• ecef91d 🤝 Improve handling of optional dependencies

• 70c6067 Merge branch 'mg/adaptive_rho_intervall'

• 64dbcc1 🚀 Scale rho-value of loose, active box constraints

• 9b50f17 Print first iteration

• 44ada54 Allow adaptive rho interval determination

• 2dcc593 Make deepcopy of settings object

• 41f82f3 🔮 Enable code literacy

• 9ea6dc2 Update newest Julia version in travis

Version 0.7.1 (11. March 2020)

• 264294d Add precompile statements (#109)
• 1b8c58c Print decomposition info only when decomposing
• d8dfe26 Prevent decomposition from mutating settings

Version 0.7.0 (05. March 2020)

• This release adds support for indirect linear system solver from the IterativeSolvers.jl package. We also drop support of PsdConeSquare constraints ins MOI. By default MOI will bridge them to PSDTriangle constraints. We further fixed a number of bugs related to the compact transformation for decomposable SDPs. Another change is that when used with verbose_timing = true, COSMO will now measure the initial factorisation time and the time for refactorisations separately. The major commits are listed here:

• 439f218 Return NearlyFeasible status code if max_iter reached

• fda29e1 Fix problem with empty columns in compact_transform

• 4324560 Drop support for MOI.PSDConeSquare

• f0a99fe ⏲ Add more detailed timing to solver

• 92aa32c Store rho-updates in result variable

• eb0be3b Assemble dual variable like in Grone's Theorem

• 14e862e Make Cholmod (SuiteSparse) default KKT solver

• e64cfc9 Fix bug with CGIndirectKKTSolver and MINRESIndirectKKTSolver

• f514caa Merge pull request #106 from oxfordcontrol/pg/pardiso_fix

• 6290cbe fixes minres

• b2a0e49 fix pardiso default options

• 5ed8a9c [ci skip] Add info about custom linear solvers

• cfa6f76 Merge pull request #92 from oxfordcontrol/cg

Version 0.6.0 (26. November 2019)

We add smart clique merging and fix a number of bugs in COSMO and the MOIWrapper.

• 94babde Add Documentation for Chordal Decomposition
• 3ba43dd Renaming merge concepts, adding docstrings
• 96e9c1c Implementation of a SparseCoLO like transformation
• 752117b Add unit test for psd completion with clique merge
• 975f87a 🔧 Fix bug about recovering the dual variable
• e524b59 Working recomputation of clique tree
• 09dc024 Implement clique merging
• f90cc62 <#96> Map var idxs when processing MOI constraints
• 9f1b5ff <#97> Loop over all MOI-Attributes during copy_to
• 59f0480 Add unit test for 1d psd projection
• ab821d8 Fix bug in 1D PSD cone projection
• 857bf18 Merge pull request #98 from rschwarz/rs/moi095
• 969cb2b Fix MOI tests "result_count".
• c3d0263 Exclude MOI unit test "number_threads".
• ed25b4b [#94] Fix bug related to DualExpCone problems
• 7f91e71 💾 Implement memory-efficient infeasibility checks
• dc75bc5 🔧 Code refactoring
• a754b0c Implement memory-efficient residual computation
• 020ada4 [#94] Merge pull request #94 from oxfordcontrol/check_psd_clean
• 4b8396d Use (faster) in-place versions for in_pos_def()
• 5565fa9 Remove sparse version of is_pos_def!()
• dceea67 Avoids eigendecomposition in psd tests.
• 340ecde Assemble entry selector matrix more efficient

Version 0.5.0 (30. August 2019)

• 8986827 Update to support MOI v0.9

Version 0.4.4 (23. August 2019)

• b514287 Fix memory issue in MOI's interface (PR #87)
• cc32a30 Remove parameter printing for sets
• d31e501 Code refactoring in chordal decomposition
• 128b409 Remove num of elem in A and size of b from header
• b267b1c Remove eigendecomposition of P in MOI wrapper
• 21eb8da Move PowerCone to typed set in MOI wrapper test

Version 0.4.3 (3. June 2019)

We add the exponential cone, the power cone and their dual cones. Furthermore, we offer the option to use different linear system solvers. The most important changes are listed here:

• eb7cb68 Add power cone and dual exp and pow cone
• f3b5c2f Add a logistic regression example
• 44cc8b3 Implement exponential cone constraints
• 4c3428c Merge pull request #65 from oxfordcontrol/pg/lin_solvers
• 863168a <#65> Support custom linear system solver

Version 0.4.2 (5. May 2019)

This is a major upgrad since it adds the chordal decomposition functionality. The most important changes are listed here:

• 8dd70eb Make sure graphs are connected
• b44364c Add chordal decomp info to header printing
• 5acaf73 Move PSDTriangle memory allocation after decomp
• 25593b4 Remove graph computation
• a115196 Add DensePsdCones to handle non-decomposable cones
• dbe70bc Add PSD completion
• 4262bfe Add chordal decomposition algorithm and unit tests
• bdebbc5 🔧 Fix bug in scaling of MOI-PSDTriangle constraint
• 726d03b <#55> Implement set merging in MOI Wrapper
• 2a3957f 🤫 Add support for MOI.Silent
• d0f0e44 Fix bugs in warm starting
• f5bd31b 32 bit BLAS support
• aeab91c Check positive semidefiniteness of P in MOIWrapper
• 9ad8f36 Fix bug in MOI termination codes
• 34ebb89 Add deepcopy step for convex sets in set!()

Version 0.4.1 (16. April 2019)

The most important changes are listed below:

• (023f517) Fix printing of rho
• (cb2a483) Memory efficient projections in BLAS
• (c582f83) Remove lower-triangle legacy code in MOI wrapper
• (6b89271) Dont export Model and convex sets
• (d98c7ea) Bug fix in norm computation
• (0ad53d6) Add Aox constraint type
• (360ca0f) Handle edge cases for P, q in assemble!

Version 0.4 (22. February 2019)

• The optional arguments settings, x0, y0, s0 are now keyword arguments (breaks backward compatibility), see d0f22aab
• Some smaller bug fixes and function renaming
• Documentation via Documenter.jl

Version 0.3.1 (2. February 2019)

• Support warm starting of primal and dual variables in MOI / JuMP interface
• Fix a bug when printing the header
• Improve unit test coverage

Version 0.3 (7. January 2019)

First public release of COSMO solver

• Add MOI / JuMP support