Skip to content

Latest commit

 

History

History
618 lines (442 loc) · 28.7 KB

NEWS.md

File metadata and controls

618 lines (442 loc) · 28.7 KB
Raw

JuMP release notes

Version 0.19 (unreleased)

JuMP 0.19 contains significant breaking changes.

We will tag alpha and beta versions of 0.19 as appropriate to make it easier to test JuMP 0.19 before the final release. These are preview releases for early adopters. Additional breaking changes may occur between development tags. The documentation has known gaps. When using a development tag, be prepared to peek into the JuMP source code and tests for examples of how things work.

Latest development tag: v0.19-alpha (] add JuMP#v0.19-alpha).

Breaking changes:

  • JuMP's abstraction layer for communicating with solvers changed from MathProgBase (MPB) to MathOptInterface (MOI). MOI addresses many longstanding design issues. (See @mlubin's slides from JuMP-dev 2018.) JuMP 0.19 is compatible only with solvers that have been updated for MOI. See the installation guide for a list of solvers that have and have not yet been updated.

  • JuMP containers (e.g., the objects returned by @variable) have been redesigned. Base.Dict replaces JuMPDict, JuMPArray was rewritten (inspired by AxisArrays), and you can now request a container type with the container= keyword to the macros. See the corresponding documentation for more details.

  • The statuses returned by solvers have changed. See the possible status values here. The MOI statuses are much richer than the MPB statuses and can be used to distinguish between previously indistinguishable cases (e.g. did the solver have a feasible solution when it stopped because of the time limit?).

  • Starting values are separate from result values. Use value to query the value of a variable in a solution. Use start_value and set_start_value to get and set an initial starting point provided to the solver. The solutions from previous solves are no longer automatically set as the starting points for the next solve.

  • The data structures for affine and quadratic expressions AffExpr and QuadExpr have changed. Internally, terms are stored in dictionaries instead of lists. Duplicate coefficients can no longer exist. Accessors and iteration methods have changed.

  • JuMPNLPEvaluator no longer includes the linear and quadratic parts of the model in the evaluation calls. These are now handled separately to allow NLP solvers that support various types of constraints.

  • JuMP solver-independent callbacks have been replaced by solver-specific callbacks. See your favorite solver for more details. (TODO: No solver-specific callbacks are implemented yet.)

  • The norm() syntax is no longer recognized inside macros. Use the SecondOrderCone() set instead.

  • JuMP no longer performs automatic transformation between special quadratic forms and second-order cone constraints. Support for these constraint classes depends on the solver.

  • The symbols :Min and :Max are no longer used as optimization senses. Instead, JuMP uses the OptimizationSense enum from MathOptInterface. @objective(model, Max, ...), @objective(model, Min, ...), @NLobjective(model, Max, ...), and @objective(model, Min, ...) remain valid, but @objective(m, :Max, ...) is no longer accepted.

  • The sign conventions for duals has changed in some cases for consistency with conic duality (see the documentation). The shadow_price helper method returns duals with signs that match conventional LP interpretations of dual values as sensitivities of the objective value to relaxations of constraints.

New features:

  • Splatting (like f(x...)) is recognized in restricted settings in nonlinear expressions.

  • Support for deleting constraints and variables.

  • The documentation has been completely rewritten using docstrings and Documenter.

  • Support for modeling mixed conic and quadratic models (e.g., conic models with quadratic objectives and bi-linear matrix inequalities).

  • Significantly improved support for modeling new types of constraints and for extending JuMP's macros.

  • Support for providing dual warm starts.

  • Improved support for accessing solver-specific attributes (e.g., the irreducible inconsistent subsystem).

  • Explicit control of whether symmetry-enforcing constraints are added to PSD constraints.

  • Support for modeling exponential cones.

  • Significant improvements in internal code quality and testing.

  • Style and naming guidelines.

  • Direct mode and manual mode provide explicit control over when copies of a model are stored and/or regenerated. See the corresponding documentation.

Known issues:

  • Model printing is not yet implemented. (issue)

  • There are known performance regressions (issue). Fast incremental NLP solves are not yet reimplemented (issue).

  • We do not yet have an implementation of solver-specific callbacks.

  • Example files (under examples/) have not yet been updated.

See the 0.19 issue milestone for a complete list of issues blocking the 0.19 release.

Version 0.18.4 (October 8, 2018)

  • Fix a bug in model printing on Julia 0.7 and 1.0.

Version 0.18.3 (October 1, 2018)

  • Add support for Julia v1.0 (Thanks @ExpandingMan)
  • Fix matrix expressions with quadratic functions (#1508)

Version 0.18.2 (June 10, 2018)

  • Fix a bug in second-order derivatives when expressions are present (#1319)
  • Fix a bug in @constraintref (#1330)

Version 0.18.1 (April 9, 2018)

  • Fix for nested tuple destructuring (#1193)
  • Preserve internal model when relaxation=true (#1209)
  • Minor bug fixes and updates for example

Version 0.18.0 (July 27, 2017)

  • Drop support for Julia 0.5.
  • Update for ForwardDiff 0.5.
  • Minor bug fixes.

Version 0.17.1 (June 9, 2017)

  • Use of constructconstraint! in @SDconstraint.
  • Minor bug fixes.

Version 0.17.0 (May 27, 2017)

  • Breaking change: Mixing quadratic and conic constraints is no longer supported.
  • Breaking change: The getvariable and getconstraint functions are replaced by indexing on the corresponding symbol. For instance, to access the variable with name x, one should now write m[:x] instead of getvariable(m, :x). As a consequence, creating a variable and constraint with the same name now triggers a warning, and accessing one of them afterwards throws an error. This change is breaking only in the latter case.
  • Addition of the getobjectivebound function that mirrors the functionality of the MathProgBase getobjbound function except that it takes into account transformations performed by JuMP.
  • Minor bug fixes.

The following changes are primarily of interest to developers of JuMP extensions:

  • The new syntax @constraint(model, expr in Cone) creates the constraint ensuring that expr is inside Cone. The Cone argument is passed to constructconstraint! which enables the call to the dispatched to an extension.
  • The @variable macro now calls constructvariable! instead of directly calling the Variable constructor. Extra arguments and keyword arguments passed to @variable are passed to constructvariable! which enables the call to be dispatched to an extension.
  • Refactor the internal function conicdata (used build the MathProgBase conic model) into smaller subfunctions to make these parts reusable by extensions.

Version 0.16.2 (March 28, 2017)

  • Minor bug fixes and printing tweaks
  • Address deprecation warnings for Julia 0.6

Version 0.16.1 (March 7, 2017)

  • Better support for AbstractArray in JuMP (Thanks @tkoolen)
  • Minor bug fixes

Version 0.16.0 (February 23, 2017)

  • Breaking change: JuMP no longer has a mechanism for selecting solvers by default (the previous mechanism was flawed and incompatible with Julia 0.6). Not specifying a solver before calling solve() will result in an error.
  • Breaking change: User-defined functions are no longer global. The first argument to JuMP.register is now a JuMP Model object within whose scope the function will be registered. Calling JuMP.register without a Model now produces an error.
  • Breaking change: Use the new JuMP.fix method to fix a variable to a value or to update the value to which a variable is fixed. Calling setvalue on a fixed variable now results in an error in order to avoid silent behavior changes. (Thanks @joaquimg)
  • Nonlinear expressions now print out similarly to linear/quadratic expressions (useful for debugging!)
  • New category keyword to @variable. Used for specifying categories of anonymous variables.
  • Compatibility with Julia 0.6-dev.
  • Minor fixes and improvements (Thanks @cossio, @ccoffrin, @blegat)

Version 0.15.1 (January 31, 2017)

  • Bugfix for @LinearConstraints and friends

Version 0.15.0 (December 22, 2016)

  • Julia 0.5.0 is the minimum required version for this release.
  • Document support for BARON solver
  • Enable info callbacks in more states than before, e.g. for recording solutions. New when argument to addinfocallback (#814, thanks @yeesian)
  • Improved support for anonymous variables. This includes new warnings for potentially confusing use of the traditional non-anonymous syntax:
    • When multiple variables in a model are given the same name
    • When non-symbols are used as names, e.g., @variable(m, x[1][1:N])
  • Improvements in iterating over JuMP containers (#836, thanks @IssamT)
  • Support for writing variable names in .lp file output (Thanks @leethargo)
  • Support for querying duals to SDP problems (Thanks @blegat)
  • The comprehension syntax with curly braces sum{}, prod{}, and norm2{} has been deprecated in favor of Julia's native comprehension syntax sum(), prod() and norm() as previously announced. (For early adopters of the new syntax, norm2() was renamed to norm() without deprecation.)
  • Unit tests rewritten to use Base.Test instead of FactCheck
  • Improved support for operations with matrices of JuMP types (Thanks @ExpandingMan)
  • The syntax to halt a solver from inside a callback has changed from throw(CallbackAbort()) to return JuMP.StopTheSolver
  • Minor bug fixes

Version 0.14.2 (December 12, 2016)

  • Allow singeton anonymous variables (includes bugfix)

Version 0.14.1 (September 12, 2016)

  • More consistent handling of states in informational callbacks, includes a new when parameter to addinfocallback for specifying in which state an informational callback should be called.

Version 0.14.0 (August 7, 2016)

  • Compatibility with Julia 0.5 and ForwardDiff 0.2
  • Support for "anonymous" variables, constraints, expressions, and parameters, e.g., x = @variable(m, [1:N]) instead of @variable(m, x[1:N])
  • Support for retrieving constraints from a model by name via getconstraint
  • @NLconstraint now returns constraint references (as expected).
  • Support for vectorized expressions within lazy constraints
  • On Julia 0.5, parse new comprehension syntax sum(x[i] for i in 1:N if isodd(i)) instead of sum{ x[i], i in 1:N; isodd(i) }. The old syntax with curly braces will be deprecated in JuMP 0.15.
  • Now possible to provide nonlinear expressions as "raw" Julia Expr objects instead of using JuMP's nonlinear macros. This input format is useful for programmatically generated expressions.
  • s/Mathematical Programming/Mathematical Optimization/
  • Support for local cuts (Thanks to @madanim, Mehdi Madani)
  • Document Xpress interface developed by @joaquimg, Joaquim Dias Garcia
  • Minor bug and deprecation fixes (Thanks @odow, @jrevels)

Version 0.13.2 (May 16, 2016)

  • Compatibility update for MathProgBase

Version 0.13.1 (May 3, 2016)

  • Fix broken deprecation for registerNLfunction.

Version 0.13.0 (April 29, 2016)

  • Most exported methods and macros have been renamed to avoid camelCase. See the list of changes here. There is a 1-1 mapping from the old names to the new, and it is safe to simply replace the names to update existing models.
  • Specify variable lower/upper bounds in @variable using the lowerbound and upperbound keyword arguments.
  • Change name printed for variable using the basename keyword argument to @variable.
  • New @variables macro allows multiline declaration of groups of variables.
  • A number of solver methods previously available only through MathProgBase are now exposed directly in JuMP. The fix was recorded live!
  • Compatibility fixes with Julia 0.5.
  • The "end" indexing syntax is no longer supported within JuMPArrays which do not use 1-based indexing until upstream issues are resolved, see here.

Version 0.12.2 (March 9, 2016)

  • Small fixes for nonlinear optimization

Version 0.12.1 (March 1, 2016)

  • Fix a regression in slicing for JuMPArrays (when not using 1-based indexing)

Version 0.12.0 (February 27, 2016)

  • The automatic differentiation functionality has been completely rewritten with a number of user-facing changes:
    • @defExpr and @defNLExpr now take the model as the first argument. The previous one-argument version of @defExpr is deprecated; all expressions should be named. E.g., replace @defExpr(2x+y) with @defExpr(jump_model, my_expr, 2x+y).
    • JuMP no longer uses Julia's variable binding rules for efficiently re-solving a sequence of nonlinear models. Instead, we have introduced nonlinear parameters. This is a breaking change, so we have added a warning message when we detect models that may depend on the old behavior.
    • Support for user-defined functions integrated within nonlinear JuMP expressions.
  • Replaced iteration over AffExpr with Number-like scalar iteration; previous iteration behavior is now available via linearterms(::AffExpr).
  • Stopping the solver via throw(CallbackAbort()) from a callback no longer triggers an exception. Instead, solve() returns UserLimit status.
  • getDual() now works for conic problems (Thanks @emreyamangil.)

Version 0.11.3 (February 4, 2016)

  • Bug-fix for problems with quadratic objectives and semidefinite constraints

Version 0.11.2 (January 14, 2016)

  • Compatibility update for Mosek

Version 0.11.1 (December 1, 2015)

  • Remove usage of @compat in tests.
  • Fix updating quadratic objectives for nonlinear models.

Version 0.11.0 (November 30, 2015)

  • Julia 0.4.0 is the minimum required version for this release.
  • Fix for scoping semantics of index variables in sum{}. Index variables no longer leak into the surrounding scope.
  • Addition of the solve(m::Model, relaxation=true) keyword argument to solve the standard continuous realaxation of model m
  • The getConstraintBounds() method allows access to the lower and upper bounds of all constraints in a (nonlinear) model.
  • Update for breaking changes in MathProgBase

Version 0.10.3 (November 20, 2015)

  • Fix a rare error when parsing quadratic expressions
  • Fix Variable() constructor with default arguments
  • Detect unrecognized keywords in solve()

Version 0.10.2 (September 28, 2015)

  • Fix for deprecation warnings

Version 0.10.1 (September 3, 2015)

  • Fixes for ambiguity warnings.
  • Fix for breaking change in precompilation syntax in Julia 0.4-pre

Version 0.10.0 (August 31, 2015)

  • Support (on Julia 0.4 and later) for conditions in indexing @defVar and @addConstraint constructs, e.g. @defVar(m, x[i=1:5,j=1:5; i+j >= 3])
  • Support for vectorized operations on Variables and expressions. See the documentation for details.
  • New getVar() method to access variables in a model by name
  • Support for semidefinite programming.
  • Dual solutions are now available for general nonlinear problems. You may call getDual on a reference object for a nonlinear constraint, and getDual on a variable object for Lagrange multipliers from active bounds.
  • Introduce warnings for two common performance traps: too many calls to getValue() on a collection of variables and use of the + operator in a loop to sum expressions.
  • Second-order cone constraints can be written directly with the norm() and norm2{} syntax.
  • Implement MathProgBase interface for querying Hessian-vector products.
  • Iteration over JuMPContainers is deprecated; instead, use the keys and values functions, and zip(keys(d),values(d)) for the old behavior.
  • @defVar returns Array{Variable,N} when each of N index sets are of the form 1:nᵢ.
  • Module precompilation: on Julia 0.4 and later, using JuMP is now much faster.

Version 0.9.3 (August 11, 2015)

  • Fixes for FactCheck testing on julia v0.4.

Version 0.9.2 (June 27, 2015)

  • Fix bug in @addConstraints.

Version 0.9.1 (April 25, 2015)

  • Fix for Julia 0.4-dev.
  • Small infrastructure improvements for extensions.

Version 0.9.0 (April 18, 2015)

  • Comparison operators for constructing constraints (e.g. 2x >= 1) have been deprecated. Instead, construct the constraints explicitly in the @addConstraint macro to add them to the model, or in the @LinearConstraint macro to create a stand-alone linear constraint instance.
  • getValue() method implemented to compute the value of a nonlinear subexpression
  • JuMP is now released under the Mozilla Public License version 2.0 (was previously LGPL). MPL is a copyleft license which is less restrictive than LGPL, especially for embedding JuMP within other applications.
  • A number of performance improvements in ReverseDiffSparse for computing derivatives.
  • MathProgBase.getsolvetime(m) now returns the solution time reported by the solver, if available. (Thanks @odow, Oscar Dowson)
  • Formatting fix for LP format output. (Thanks @sbebo, Leonardo Taccari).

Version 0.8.0 (February 17, 2015)

  • Nonlinear subexpressions now supported with the @defNLExpr macro.
  • SCS supported for solving second-order conic problems.
  • setXXXCallback family deprecated in favor of addXXXCallback.
  • Multiple callbacks of the same type can be registered.
  • Added support for informational callbacks via addInfoCallback.
  • A CallbackAbort exception can be thrown from callback to safely exit optimization.

Version 0.7.4 (February 4, 2015)

  • Reduced costs and linear constraint duals are now accessible when quadratic constraints are present.
  • Two-sided nonlinear constraints are supported.
  • Methods for accessing the number of variables and constraints in a model are renamed.
  • New default procedure for setting initial values in nonlinear optimization: project zero onto the variable bounds.
  • Small bug fixes.

Version 0.7.3 (January 14, 2015)

  • Fix a method ambiguity conflict with Compose.jl (cosmetic fix)

Version 0.7.2 (January 9, 2015)

  • Fix a bug in sum(::JuMPDict)
  • Added the setCategory function to change a variables category (e.g. continuous or binary) after construction, and getCategory to retrieve the variable category.

Version 0.7.1 (January 2, 2015)

  • Fix a bug in parsing linear expressions in macros. Affects only Julia 0.4 and later.

Version 0.7.0 (December 29, 2014)

Linear/quadratic/conic programming

  • Breaking change: The syntax for column-wise model generation has been changed to use keyword arguments in @defVar.
  • On Julia 0.4 and later, variables and coefficients may be multiplied in any order within macros. That is, variable*coefficient is now valid syntax.
  • ECOS supported for solving second-order conic problems.

Nonlinear programming

  • Support for skipping model generation when solving a sequence of nonlinear models with changing data.
  • Fix a memory leak when solving a sequence of nonlinear models.
  • The @addNLConstraint macro now supports the three-argument version to define sets of nonlinear constraints.
  • KNITRO supported as a nonlinear solver.
  • Speed improvements for model generation.
  • The @addNLConstraints macro supports adding multiple (groups of) constraints at once. Syntax is similar to @addConstraints.
  • Discrete variables allowed in nonlinear problems for solvers which support them (currently only KNITRO).

General

  • Starting values for variables may now be specified with @defVar(m, x, start=value).
  • The setSolver function allows users to change the solver subsequent to model creation.
  • Support for "fixed" variables via the @defVar(m, x == 1) syntax.
  • Unit tests rewritten to use FactCheck.jl, improved testing across solvers.

Version 0.6.3 (October 19, 2014)

  • Fix a bug in multiplying two AffExpr objects.

Version 0.6.2 (October 11, 2014)

  • Further improvements and bug fixes for printing.
  • Fixed a bug in @defExpr.
  • Support for accessing expression graphs through the MathProgBase NLP interface.

Version 0.6.1 (September 19, 2014)

  • Improvements and bug fixes for printing.

Version 0.6.0 (September 9, 2014)

  • Julia 0.3.0 is the minimum required version for this release.
  • buildInternalModel(m::Model) added to build solver-level model in memory without optimizing.
  • Deprecate load_model_only keyword argument to solve.
  • Add groups of constraints with @addConstraints macro.
  • Unicode operators now supported, including ∑ for sum, ∏ for prod, and ≤/≥
  • Quadratic constraints supported in @addConstraint macro.
  • Quadratic objectives supported in @setObjective macro.
  • MathProgBase solver-independent interface replaces Ipopt-specific interface for nonlinear problems
    • Breaking change: IpoptOptions no longer supported to specify solver options, use m = Model(solver=IpoptSolver(options...)) instead.
  • New solver interfaces: ECOS, NLopt, and nonlinear support for MOSEK
  • New option to control whether the lazy constraint callback is executed at each node in the B&B tree or just when feasible solutions are found
  • Add support for semicontinuous and semi-integer variables for those solvers that support them.
  • Add support for index dependencies (e.g. triangular indexing) in @defVar, @addConstraint, and @defExpr (e.g. @defVar(m, x[i=1:10,j=i:10])).
    • This required some changes to the internal structure of JuMP containers, which may break code that explicitly stored JuMPDict objects.

Version 0.5.8 (September 24, 2014)

  • Fix a bug with specifying solvers (affects Julia 0.2 only)

Version 0.5.7 (September 5, 2014)

  • Fix a bug in printing models

Version 0.5.6 (September 2, 2014)

  • Add support for semicontinuous and semi-integer variables for those solvers that support them.
    • Breaking change: Syntax for Variable() constructor has changed (use of this interface remains discouraged)
  • Update for breaking changes in MathProgBase

Version 0.5.5 (July 6, 2014)

  • Fix bug with problem modification: adding variables that did not appear in existing constraints or objective.

Version 0.5.4 (June 19, 2014)

  • Update for breaking change in MathProgBase which reduces loading times for using JuMP
  • Fix error when MIPs not solved to optimality

Version 0.5.3 (May 21, 2014)

  • Update for breaking change in ReverseDiffSparse

Version 0.5.2 (May 9, 2014)

  • Fix compatibility with Julia 0.3 prerelease

Version 0.5.1 (May 5, 2014)

  • Fix a bug in coefficient handling inside lazy constraints and user cuts

Version 0.5.0 (May 2, 2014)

  • Support for nonlinear optimization with exact, sparse second-order derivatives automatically computed. Ipopt is currently the only solver supported.
  • getValue for AffExpr and QuadExpr
  • Breaking change: getSolverModel replaced by getInternalModel, which returns the internal MathProgBase-level model
  • Groups of constraints can be specified with @addConstraint (see documentation for details). This is not a breaking change.
  • dot(::JuMPDict{Variable},::JuMPDict{Variable}) now returns the corresponding quadratic expression.

Version 0.4.1 (March 24, 2014)

  • Fix bug where change in objective sense was ignored when re-solving a model.
  • Fix issue with handling zero coefficients in AffExpr.

Version 0.4.0 (March 10, 2014)

  • Support for SOS1 and SOS2 constraints.
  • Solver-independent callback for user heuristics.
  • dot and sum implemented for JuMPDict objects. Now you can say @addConstraint(m, dot(a,x) <= b).
  • Developers: support for extensions to JuMP. See definition of Model in src/JuMP.jl for more details.
  • Option to construct the low-level model before optimizing.

Version 0.3.2 (February 17, 2014)

  • Improved model printing
    • Preliminary support for IJulia output

Version 0.3.1 (January 30, 2014)

  • Documentation updates
    • Support for MOSEK
    • CPLEXLink renamed to CPLEX

Version 0.3.0 (January 21, 2014)

  • Unbounded/infeasibility rays: getValue() will return the corresponding components of an unbounded ray when a model is unbounded, if supported by the selected solver. getDual() will return an infeasibility ray (Farkas proof) if a model is infeasible and the selected solver supports this feature.
  • Solver-independent callbacks for user generated cuts.
  • Use new interface for solver-independent QCQP.
  • setlazycallback renamed to setLazyCallback for consistency.

Version 0.2.0 (December 15, 2013)

  • Breaking change: Objective sense is specified in setObjective instead of in the Model constructor.
  • Breaking change: lpsolver and mipsolver merged into single solver option.
  • Problem modification with efficient LP restarts and MIP warm-starts.
  • Relatedly, column-wise modeling now supported.
  • Solver-independent callbacks supported. Currently we support only a "lazy constraint" callback, which works with Gurobi, CPLEX, and GLPK. More callbacks coming soon.

Version 0.1.2 (November 16, 2013)

  • Bug fixes for printing, improved error messages.
  • Allow AffExpr to be used in macros; e.g., ex = y + z; @addConstraint(m, x + 2*ex <= 3)

Version 0.1.1 (October 23, 2013)

  • Update for solver specification API changes in MathProgBase.

Version 0.1.0 (October 3, 2013)

  • Initial public release.