CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog (since argmin version 0.6.0), and this project adheres to Semantic Versioning (since argmin version 0.5.0).

argmin unreleased

• Added a new GUI observer called Spectator (argmin-observer-spectator and spectator packages) (@stefan-k, #311)

argmin-math unreleased

argmin v0.10.0 2024-02-27

Added

• Added optional timeout to Executor (@stefan-k, #405)

Changed

• Users can now set ParticleSwarms random number generator (@jonboh, #383)
• Moved all observers and checkpointing into dedicated crates. This led to substantially fewer dependencies for argmin itself
  • SlogLogger moved into argmin-observer-slog (@stefan-k, #311)
  • WriteToFile was renamed to ParamWriter and moved into argmin-observer-paramwriter (@stefan-k, #395)
  • FileCheckpoint moved into argmin-checkpointing-file (@stefan-k, #311)
• Added state to observe_init of Observe trait (@stefan-k, #311)
• All crates are now in the crates directory (@stefan-k, #415)
• All examples are now crates and are in the examples directory. (@stefan-k, #387)
• Removed SerializeAlias and DeserializeAlias traits since they are not necessary anymore (@stefan-k, #395)
• Residuals in CG are now handled by IterState (@stefan-k, #408)
• Interrupt handling now includes SIGINT, SIGTERM and SIGHUP and KeyboardInterrupt was renamed to Interrupt (@stefan-k, #413)
• Error handling in LBFGS was slightly improved (@stefan-k, #416)
• KvValue::get_float(&self) now works for all kinds of KvValue (@stefan-k, #311)
• Updated gnuplot from 0.0.39 to 0.0.4 (#394)

Fixed

• Fixed residuals handling in GaussNewton (@gmilleramilar, @stefan-k, #343, #392)

argmin-math v0.4.0 2024-02-27

Added

• Implemented ArgminInv for f32/f64 (1D) (@sdrap, #346)
• Test cases for ArgminMinMax were added (@Shreyan11, #391)

Changed

• Default linalg backend for development is now MKL, which makes development on Windows possible. This required all ndarray tests to be moved in dedicated crates (@Tastaturtaste, #369)
• ArgminRandom::rand_from_range(..) now also takes a random number generator. This allows for setting the seed manually (@jonboh, #383)

Removed

• All *-serde features were removed, as they are not necessary anymore. (@stefan-k, #387)

argmin-observer-slog v0.1.0 2024-02-27

Initial release.

argmin-observer-paramwriter v0.1.0 2024-02-27

Initial release.

argmin-checkpointing-file v0.1.0 2024-02-27

Initial release.

argmin-testfunctions-py v0.0.1 2024-02-16

• The first version of a Python wrapper around argmin_testfunctions was released.
• The fact that this is still experimental is expressed with the version number 0.0.1. Eventually this module's version may end up in lockstep with argmin_testfunctions' version number.

argmin_testfunctions v0.2.0 2024-02-16

All work of this release was done by @stefan-k.

Added

• Added derivative and Hessian calculation to all test functions.
• Tested all derivatives and Hessian against finite differences (with finitediff) on multiple positions with proptest.
• Added const generics versions for the derivatives and Hessians of multi dimensional test functions to avoid allocations.
• Those test functions which operate on a fixed number of parameters now take a fixed length array as input instead of a slice.

Fixed

• Bugs in the Levy test function were fixed.

Changed

• Rust edition 2021
• Repository was moved into argmin monorepo
• rosenbrock now only takes the parameters as input. For setting the additional optional parameters a and b one can use rosenbrock_ab
• All rosenbrock_2d* functions were removed.

argmin v0.9.0 2024-01-06

Added

• Added a simple example of Nelder-Mead usage (@cjordan, #359)

Fixed

• Clippy lints (@jjbayer, @stefan-k, #341, #376, #380)
• Fixed wrong example link (@itrumper, #352)
• Correct line search to search over gradient (@DevonMorris, #354)
• SteepestDescent now correctly keeps track of prev_param (@DevonMorris, #362)
• Added a missing feature for testing argmin-math in the docs (@stefan-k, #371)
• Fixed a typo in argmin book (@imeckler, #373)
• Fixed crate versions in book (@stefan-k, #375)

Changed

• Updated gnuplot to 0.0.39 (@stefan-k, #364)
• Switched to resolver = 2 for entire workspace (@stefan-k, #372)

argmin v0.8.1 2023-02-20

Added

• Mentioned cobyla in docs and README (@relf)

Fixed

• Restored Serialize/Deserialize derives for ObserverMode (@maoe)

Changed

• Updated ctrlc to 3.2.4 (@stefan-k)
• Updated rayon to 1.6 (@stefan-k)
• Updated slog to 2.7 (@stefan-k)
• Updated slog-term to 2.9 (@stefan-k)
• Updated slog-json to 2.6 (@stefan-k)

argmin v0.8.0 2023-01-28

Added

• Added a full feature which enables all features (@stefan-k)
• Added a particle swarm optimization example using the nalgebra math backend (@stefan-k)
• Added a L-BFGS example using the nalgebra math backend (@stefan-k)

Changed

• Improved termination handling (@relf, #305)
  • Introduced TerminationStatus which can either be Terminated(TerminationReason) or NotTerminated
  • The fields of the TerminationReason enum were reduced to cases potentially relevant to any solver. Solver-specific cases can be handled with SolverExit(String).
  • Added new TerminationReason::KeyboardInterrupt and TerminationReason::Converged
• Changes to KV store used to get values from the solver to observers (@stefan-k)
  • KV store is now typed (via KvValue), which means that observers can now use the actual numbers (in the past those values were only dyn Display and as such could only be printed)
  • make_kv! renamed to kv!
• Solver does not need to be Serialize anymore when serde1 feature is enabled. This was an oversight reported by @relf and fixed by @stefan-k
• Better calculation of elapsed time inside Executor (@TheIronBorn)
• The check whether the target cost is reached is now based on the current best cost function value rather than the current cost function value (@relf)

argmin-math v0.3.0 2023-01-28

Added

• Added support for nalgebra 0.32

Changed

• All math backends now implement all math related traits (apart from the Vec backend which does not implement ArgminInv). Therefore (almost) all solvers work with all backends. (@hypotrochoid, @stefan-k)
• More primitive and complex types are now covered as well (@stefan-k)
• Reached 100% test coverage (@stefan-k)
• InverseError is not Clone anymore (@stefan-k)
• Upgraded ndarray-linalg to version 0.16 for the ndarray 0.15 backend

argmin-math v0.2.1 2022-09-09

Added

• Added ArgminMinMax and ArgminSignum implementation for nalgebra types. Now the L-BFGS should work work with the nalgebra backend again. (#257, #258, @stefan-k)
• Added an L-BFGS example using the nalgebra backend (#258, @stefan-k)

argmin v0.7.0 and argmin-math v0.2.0 2022-08-28

Added

• L1 regularization added to L-BFGS (also known as OWL-QN). L-BFGS now has more trait bounds, please consult the documentation if this causes problems. (#244, #248, #250, Thanks to @vbkaisetsu)

Changed

• The ArgminL1Norm trait was added by @vbkaisetsu for the L1 regularization in L-BFGS. In order to make the L2 norm more consistent, the corresponding trait was renamed from ArgminNorm to ArgminL2Norm and the norm method was renamed to l2_norm. (#253, @stefan-k)

Fixed

• Version 0.6 accidentally removed the possibility to deactivate the ndarray-linalg dependency in argmin-math. This leads to problems in downstream crates which do not need the functionality of ndarray-linalg and want to avoid linking angainst a BLAS. In this version, this functionality was added again. Please consult the documentation of argmin-math for details on the available backends and their configuration. (#243, #249, @stefan-k)
• Fixed type alias names in examples (#245, Thanks to @vbkaisetsu)
• Fixed a bug in dogleg method (#246, #247, @stefan-k, Thanks to @renato145 for reporting!)

argmin v0.6.0 and argmin-math v0.1.0 2022-08-09

This is a rather large release with many (breaking) changes.

• Code related to the abstraction of math backends was split of into the argmin-math crate. This crate offers now the traits related to math operations as well as their implementations for various backends (vec, ndarray, nalgebra). The backends can be turned on and off as needed. It also supports different versions of the backends.
• The serde dependency is now optional and feature-gated via serde1. Disabling the serde1 feature disables checkpointing and some logging.
• The ArgminOp trait was removed and replaced with more specialized traits such as Operator, CostFunction, Gradient, Jacobian, Hessian, Anneal. All of the above mentioned traits come with bulk_* methods which allow one to compute cost function, gradients,... in parallel as long as the rayon feature is enabled. This is so far used in Particle Swarm Optimization.
• The handling of solver-internal state was redesigned completely to allow for various kinds of state, depending on the solver. There is a State trait now which defines the basic functionality that a state must offer.
• Many aspects were redesigned to avoid unnecessary cloning
• ArgminResult was renamed to OptimizationResult and also has a more useful API now. It returns the solver, the problem and the final internal state.
• The Solver trait was redesigned
• The solver is now configured/initialized via the configure method on Executor
• KV does not store Strings anymore but rather dyn Display. This avoids unnecessarily stringifying variables when no observer is used.
• Checkpointing is now a trait (so everyone can implement their own checkpointing mechanisms). Also the handling of checkpoints was completely redesigned.
• Made the API surface a bit more consistent.
• The prelude was removed
• OpWrapper was renamed to Problem and has now a more consistent interface. It still keeps track of the function evaluations, but to be flexible enough to support future or user-defined traits, these are stored in a HashMap now.
• PSO was redesigned and a PopulationState was introduced which replaces IterState in PSO
• Unnecessary trait bounds on solvers were removed
• Removed many unwraps and instead return errors with meaningful error messages (for instance when something isn't initialized properly)
• The optional stdweb dependency was removed
• Documentation was improved. The README is now rather short and not created from the lib.rs file anymore.
• The website was updated and the "argmin book" was published, which is so far mainly a collection of short tutorials but will be extended in the future.
• A lot more tests and doc tests were added
• ArgminError is now non_exhaustive.
• Added support for nalgebra 0.31
• Many now useless macros were removed
• Improved the CI to better test all aspects of the code
• Renamed termination_reaosn method of State to terminate_with
• Renamed all *grad* methods of IterState to *gradient*
• Code of conduct was added
• Brent's optimization method was added (#77, thanks to @Armavica)
• Fixed a bug in Nelder-Mead which was reported by @bivhitscar (thanks!)
• Remove the Default trait bound on TrustRegion and Steihaug impls (#192, thanks to @cfunky)

argmin v0.5.1 (16 February 2022)

• Fixed Bug in HagerZhang line search (#2, #184, @wfsteiner)
• Removed Default trait bounds on TrustRegion and Steihaug implementations (#187, #192, @cfunky)
• Inverse Hessians are now part of IterState, therefore the final inverse Hessian can be retrieved after an optimization run (#185, #186, @stefan-k)

argmin v0.5.0 (10 January 2022)

• Faster CI pipeline (#179, @stefan-k)
• Removed CircleCI and added rustfmt check to Github Actions (#178, @stefan-k)
• Automatically build documentation in CI when merging a PR into main (#149, #176, @stefan-k)
• Added a section to documentation where to find examples for current release and main branch, removed other links (#145, #174, @stefan-k)
• Fixed warnings when building docs and added building docs to the CI (#173, @stefan-k)
• The required features for each example are now indicated in Cargo.toml (#171, #147, @stefan-k)
• CI now includes compiling to various WASM targets (#89, #170, @stefan-k)
• Branch master renamed to main (#148, @stefan-k)
• nalgebra updated from 0.29.0 to 0.30.0 (#169)
• WASM features now mentioned in documentation and README.md (#167, @Glitchy-Tozier)
• Added tests for backtracking linesearch (#168, @stefan-k)
• Removed unsafe code from the vec-based math module (#166, @stefan-k)
• Added tests for GaussNewton method and fixed GaussNewton example (#164, @stefan-k)
• Added tests for Newton method (#163, @stefan-k)
• Treat a new parameter as "best" when both current and previous cost function values are Inf (#162, @stefan-k)
• Corrected documentation of PSO and removed an unnecessary trait bound on Hessian (#161, #141, @stefan-k, @TheIronBorn)
• Moved to edition 2021 (#160, @stefan-k)
• SA acceptance now based on current cost, not previous (fix) (#157, #159, @stefan-k, @TheIronBorn)
• LineSearchCondition now uses references (#158, @w1th0utnam3)
• Counting of sub problem function counts fixed (#154, #156, @stefan-k, @w1th0utnam3)
• Fixed incorrect checking for new best solution (#151, #152, @stefan-k, @Glitchy-Tozier)
• Fixed simulated annealing always accepting the first iteration (#150, #153, @dariogoetz)
• Fixed inconsistency between state and alpha value in Backtracking linesearch (#155, @w1th0utnam3)
• Improved clippy linting in CI (#146, @stefan-k)
• Unnecessary semi-colon in macro removed (#143, @CattleProdigy)
• Allow any RNG in SA and improve example (#139, @TheIronBorn)
• Make use of slog a feature, improve tests (#136, @ThatGeoGuy)

argmin v0.4.7 (14 August 2021)

• Moved to Github actions (#130, @stefan-k)
• Updated README.md (#131, @stefan-k)
• Updated nalgebra from 0.28 to 0.29 (#133)

argmin v0.4.6 (18 July 2021)

• updated dependencies (#121, #123, #129, @stefan-k):
  • ndarray 0.15
  • ndarray-linalg 0.14
  • appox 0.5
  • nalgebra 0.28
  • ndarray-rand 0.14
  • num-complex 0.4
  • finitediff 0.1.4

argmin v0.4.5

• Squash warnings for Nalgebra 0.26.x (#118, #117, @CattleProdigy)

argmin v0.4.4

• Finally started writing a changelog.
• Performance improvements (#111, #112, @sdd)

argmin v0.4.3

• Downgraded argmin-rand to 0.13 to match ndarray 0.14

argmin v0.4.2

• Fix lazy evaluation of gradients in line searches (#101, @w1th0utnam3)
• Various updated dependencies

argmin v0.4.1

• Typo

argmin 0.4.0

• nalgebra support (#68, @Maher4Ever)
• remove unnecessary Default bound on NelderMead (#73, @vadixidav)
• Various updated dependencies

argmin 0.3.1

• remove finitediff from ndarrayl feature (#61, @optozorax)
• MoreThuente: Added error check for NaN or Inf (#57, @MattBurn)

argmin 0.3.0

• Golden-section search (#49, @nilgoyette)
• Allow users to choose floating point precision (#50, @stefan-k)
• Remove Clone trait bound from ArgminOp (#48, @stefan-k)
• Remove Serialize trait bound on ArgminOp (#36, @stefan-k)
• Moved from failure to anyhow and thiserror (#44, @stefan-k)
• Added easier access to op and state of ArgminResult (#45, @stefan-k)
• No reexport of argmin_testfunctions (#46, @stefan-k)
• Exposed stopping criterion tolerances of Quasi-Newton methods to user(#43, @stefan-k)
• Exposed stopping criterion tolerance of NewtonCG method to user (@stefan-k)
• Exposed stopping criterion tolerances of Gauss Newton methods to user (@stefan-k)
• Exposed L-BFGS stopping criterion tolerances to user (#37, @stefan-k)
• Removed need for unwrap in cstep MoreThuente LineSearch (#38, @MattBurn)
• Removed Send and Sync trait bounds from ArgminOp (#33, @stefan-k)

argmin 0.2.6

• Brent's method (#22, @xemwebe)

argmin 0.2.5

• Particle Swarm Optimization (@jjbayer)
• Derive Clone trait (#14, @rth)
• Test convergence (#13, @rth)
• Lints (#11, @rth)
• Improvements in CG method (@stefan-k)

argmin 0.2.4

• CG improvements (@stefan-k)

older versions

For older versions please see the Git history.