Changelog

Versions 2.0.0 and 2.0.1 (2024-12-06)

  • Support Python 3.12 and 3.13

  • Drop support for macOS 11 (not supported by GitHub Actions anymore)

  • Drop support for macOS 12 (not supported by GitHub Actions anymore)

  • Fix the universal2 wheels for macOS to actually run on M1 Macs

  • Major Implement single-peaked criteria:

    • on the command-line, lincs generate classification-problem has a new --allow-single-peaked-criteria option

    • in the Problem file format, there is a new value single-peaked for preference_direction``

    • in the Model file format, there is a new possible value for accepted_values.kind: intervals, used with intervals: [[20, 80], [40, 60]]

    • in the Python API:

      • there is a new value in Criterion.PreferenceDirection: single_peaked, typically used with lc.generate_problem(..., allowed_preference_directions=[..., lc.Criterion.PreferenceDirection.single_peaked])

      • there is a new value in AcceptedValues.Kind: intervals

      • some strategies for LearnMrsortByWeightsProfilesBreed must explicitly opt-in to support single-peaked criteria by calling their base class constructor with supports_single_peaked_criteria=True

      • Breaking Adapt parts of the Python API to support single-peaked criteria

    • all the learning approaches work

    • the human-readable output of lincs describe has changed slightly to accommodate for single-peaked criteria

    • lincs visualize fails when called with single-peaked criteria. See this discussion

    • documented in our “Single-peaked criteria” guide

  • Breaking Let some thresholds be unreachable (Fix bug found in real-life ASA data)

    • in the Model file format, the last items in a thresholds or intervals list can be null

    • in the Python API, the last items in a thresholds or intervals list can be None

    • (there is a runtime check that all are null after the first null)

  • Breaking Split LearnMrsortByWeightsProfilesBreed.LearningData into PreprocessedLearningSet and LearnMrsortByWeightsProfilesBreed.ModelsBeingLearned

  • Breaking Rename LearningData.urbgs to ModelsBeingLearned.random_generators

  • Changes behavior slightly Optimize WPB model weights before evaluating and returning them

  • Breaking Add parameter weights_optimization_strategy to the ReinitializeLeastAccurate breeding strategy (required to fix previous point)

  • Changes behavior slightly Always keep the best model during WPB

  • Changes behavior slightly Upgrade to OR-Tools 9.11 and CUDA 12.4

  • Add method .check_consistency_with(problem: Problem) to Model and Alternatives

  • Add methods ModelsBeingLearned.get_model and .recompute_accuracy

Version 1.1.0 (2024-02-09)

Publish the Python API

This release establishes the second tier of the stable API: the Python interface.

Improve lincs visualize

  • Replace legend by colored text annotations

  • Support visualizing single-criterion models

  • Add one graduated vertical axis per criterion

See the following “before” and “after” images:

_images/model-1.0.png _images/model-1.1.png

Smaller changes

  • Breaking: Drop support for Python 3.7

  • Support discrete criteria (with enumerated or integer values)

  • Homogenize post-processing: this changes the numerical values of the thresholds learned by SAT-based approaches, but not their achieved accuracy

  • Improve validation of input data (e.g. check consistency between problem, model, and alternatives)

  • Build a Docker image on release (not published since 0.5.0), documented in our “Get started” guide

  • Support copy.deepcopy on I/O objects

  • Support pickling of I/O objects

  • Fix TerminateAfterIterations and TerminateAfterIterationsWithoutProgress strategies: they were allowing slightly too many iterations

  • Expose parameters of EvalMaxSAT in our API and command-line interface (see lincs learn classification-model --help):

    • --ucncs.max-sat-by-separation.solver (for consistency, always "eval-max-sat" for now)

    • --ucncs.max-sat-by-separation.eval-max-sat.nb-minimize-threads

    • --ucncs.max-sat-by-separation.eval-max-sat.timeout-fast-minimize

    • --ucncs.max-sat-by-separation.eval-max-sat.coef-minimize-time

    • --ucncs.max-sat-by-coalitions.solver (for consistency, always "eval-max-sat" for now)

    • --ucncs.max-sat-by-coalitions.eval-max-sat.nb-minimize-threads

    • --ucncs.max-sat-by-coalitions.eval-max-sat.timeout-fast-minimize

    • --ucncs.max-sat-by-coalitions.eval-max-sat.coef-minimize-time

Version 1.0.0 (2023-11-22)

This is the first stable release of lincs. It establishes the first tier of the stable API: the command-line interface.

  • Add a roadmap in the documentation

Version 0.11.1

This is the third release candidate for version 1.0.0.

  • Technical refactoring

Version 0.11.0

This is the second release candidate for version 1.0.0.

  • Breaking Rename category_correlation to preference_direction in problem files

  • Breaking Rename the growing preference direction to increasing in problem files

  • Breaking Rename the categories attribute in problem files to ordered_categories in problem files

  • Make names of generated categories more explicit (“Worst category”, “Intermediate category N”, “Best category”)

  • Support isotone (resp. antitone) as a synonym for increasing (resp. decreasing) in problem files

  • Add lincs describe command to produce human-readable descriptions of problems and models

  • Remove comments about termination conditions from learned models, but:

  • Add --mrsort.weights-profiles-breed.output-metadata to generate in YAML the data previously found in those comments

  • Provide a Jupyter notebook to help follow the “Get Started” guide (and use Jupyter for all integration tests)

  • Document the “externally managed” error on Ubuntu 23.4+

(In versions below, the term “category correlation” was used instead of “preference direction”.)

Versions 0.10.0 to 0.10.3

This is the first release candidate for version 1.0.0.

  • Breaking: Allow more flexible description of accepted values in the model json schema. See user guide for details.

  • Breaking: Rename option --ucncs.approach to --ucncs.strategy

  • Breaking: Rename option --output-classified-alternatives to --output-alternatives

  • Fix line ends on Windows

  • Fix lincs visualize to use criteria’s min/max values and category correlation

  • Validate consistency with problem when loading alternatives or model files

  • Output “reproduction command” in lincs classify

  • Improve documentation

Versions 0.9.0 to 0.9.2

  • Pre-process the learning set before all learning algorithms.

Possible values for each criterion are listed and sorted before the actual learning starts so that learning algorithms now see all criteria as:

  • having increasing correlation with the categories

  • having values in a range of integers

This is a simplification for implementers of learning algorithms, and improves the performance of the weights-profiles-breed approach.

  • Expose SufficientCoalitions::upset_roots to Python

  • Fix alternative names when using the --max-imbalance option of lincs generate classified-alternatives

  • Produce cleaner error when --max-imbalance is too tight

  • Print number of iterations at the end of WPB learnings

  • Display lincs’ version in the “Reproduction command” comment in generated files

  • Various improvements to the code’s readability

Version 0.8.7

  • Integrate CUDA parts on Windows

  • Compile with OpenMP on Windows

Versions 0.8.5 to 0.8.6

  • Distribute binary wheels for Windows!

Versions 0.8.0 to 0.8.4

  • Rename option --...max-duration-seconds to --...max-duration

  • Display termination condition after learning using the weights-profiles-breed approach

  • Make termination of the weights-profiles-breed approach more consistent

  • Integrate Chrones (as an optional dependency, on Linux only)

  • Display iterations in --...verbose mode

  • Fix pernicious memory bug

Version 0.7.0

Bugfixes:

  • Fix the Linux wheels: make sure they are built with GPU support

  • Improve building lincs without nvcc (e.g. on macOS):

    • provide the lincs info has-gpu command

    • adapt lincs learn classification-model --help

Features:

  • Add “max-SAT by coalitions” and “max-SAT by separation” learning approaches (hopefully correct this time!)

  • Use YAML anchors and aliases to limit repetitions in the model file format when describing \(U^c \textsf{-} NCS\) models

  • Specifying the minimum and maximum values for each criterion in the problem file:

    • Generate synthetic data using these attributes (--denormalized-min-max)

    • Adapt the learning algorithms to use these attributes

  • Support criteria with decreasing correlation with the categories:

    • in the problem file

    • when generating synthetic data (--allow-decreasing-criteria)

    • in the learning algorithms

  • Add a comment to all generated files stating the command-line to use to re-generate them

  • Use enough decimals when storing floating point values in models to avoid any loss of precision

  • Log final accuracy with --mrsort.weights-profiles-breed.verbose

  • Improve tests

Version 0.6.0

  • Remove buggy “max-SAT by coalitions” approach

  • Add “SAT by separation” approach

Version 0.5.1

  • Publish wheels for macOS

Version 0.5.0

  • Implement “SAT by coalitions” and “max-SAT by coalitions” removed in 0.6.0 learning methods

  • Add misclassify_alternatives to synthesize noise on alternatives

  • Expend the model file format to support specifying the sufficient coalitions by their roots

  • Produce “manylinux_2_31” binary wheels

  • Improve YAML schemas for problem and model file formats

  • Use the “flow” YAML formatting for arrays of scalars

  • Improve consistency between Python and C++ APIs (not yet documented though)

  • Add more control over the “weights, profiles, breed” learning method (termination strategies, “verbose” option)

  • Add an expansion point for the breeding part of “weights, profiles, breed”

  • Add an exception for failed learnings

Version 0.4.5

  • Use JSON schemas to document and validate the problem and model files

  • Support development on macOS and on machines without a GPU

  • Improve documentation

Versions 0.4.1 to 0.4.4

Never properly published

Version 0.4.0

  • Add a GPU (CUDA) implementation of the accuracy heuristic strategy for the “weights, profiles, breed” learning method

  • Introduce Alglib as a LP solver for the “weights, profiles, breed” learning method

  • Publish a Docker image with lincs installed

  • Change “domain” to “problem” everywhere

  • Improve documentation

  • Improve model and alternatives visualization

  • Expose ‘Alternative::category’ properly

Versions 0.3.4 to 0.3.7

  • Improve documentation

Version 0.3.3

  • Fix Python package description

Version 0.3.2

  • License (LGPLv3)

Version 0.3.1

  • Fix installation (missing C++ header file)

Version 0.3.0

  • Implement learning an MR-Sort model using Sobrie’s heuristic on CPU

Version 0.2.2

  • Add options: generate model –mrsort.fixed-weights-sum and generate classified-alternatives –max-imbalance

Version 0.2.1

  • Fix images on the PyPI website

Version 0.2.0

  • Implement generation of pseudo-random synthetic data

  • Implement classification by MR-Sort models

  • Kick-off the documentation effort with a quite nice first iteration of the README

Version 0.1.3

Initial publication with little functionality