Skip to content

CHANGES.mp.md

Latest commit

 

History

History
730 lines (605 loc) · 27.5 KB

CHANGES.mp.md

File metadata and controls

730 lines (605 loc) · 27.5 KB
Raw

Summary of recent updates to the AMPL MP Library

Unreleased

  • Fix undefined behavior in shared constraint/expression acceptance queries: no-argument calls no longer bind a null pointer to a model API reference. Existing virtual API-taking calls and acceptance levels are preserved.

20261029

  • Fix: with NL output (MP2NL), complementarity constraints were reformulated (by cvt:compl) instead of being passed as NL complementarity, and quadratic constraints were linearized, since 20261029. Quadratic constraints are linearized for NL output only with cvt:quadcon=0, also when not accepted otherwise (before, the NL output ignored cvt:quadcon=0).
  • New: AMPL's floor, ceil, round(x[, n]) (ties to even), trunc(x[, n]), div and mod (also for real numbers and variable divisors), before rejected as "unsupported". MIP reformulations: floor/ceil by bounds with cvt:mip:eps, round with an even/odd decomposition, div/mod by quotient and remainder as in MiniZinc's linearization library. Options acc:floor, acc:ceil, acc:round, acc:intdiv, acc:mod.
  • Products of variables with a bounded integer factor are reformulated exactly for MIP solvers without (nonconvex) quadratics, by binary expansion of the integer factor (option cvt:mip:intprodbits, default 30; 0 restores the piecewise-linear approximation); integer powers x^k likewise. Before, such products were approximated.
  • Fix: with cvt:quadcon=0, quadratic range constraints were still passed to solvers accepting them (e.g., MOSEK).
  • New functional constraint r = x*y (acc:mul) for solvers taking products of variables natively (CP-SAT): quadratic terms of non-binary variables become such constraints instead of the piecewise-linear approximation of ((x+y)^2-x^2-y^2)/2; equal products share one constraint. The default of cvt:multoutcard is now per solver (1 for CP-SAT: a product of sums stays one product of two auxiliary variables).
  • Fix: inlining algebraic subexpressions into an objective (cvt:pre:unnest) could make it quadratic for solvers without quadratic objectives (cvt:quadobj=0), which then failed.
  • Fix: checking whether an auxiliary variable can be replaced by its defining expression (e.g., for MP2NL and expression-based solvers) could add an unused copy of a comparison to the model.
  • Fix: a dual ray requested for a model without constraints (MOSEK, "infeasible or unbounded") failed with "map::at".
  • Fix: tech:writegraph wrote empty names for auxiliary variables.
  • Fix: a nonconvex piecewise-linear term whose argument becomes fixed at a breakpoint failed with "MIPFlatConverter: vector" in solvers without native SOS2 (e.g., HiGHS, MOSEK, or acc:sos2=0). The term now reduces to its value at that point. A term with a single linear piece over its argument's domain becomes one linear equality, without an SOS2 constraint.
  • Fix: chained piecewise-linear terms (one term's argument depends on another term's result) could get wrong solutions in solvers without native SOS2. The results of piecewise-linear terms now receive bounds from their arguments' domains, also when these are narrowed later, e.g., by linear constraints.
  • Change: the MIP reformulation of a nonconvex piecewise-linear term (solvers without native SOS2, e.g., HiGHS, MOSEK, or acc:pl=0 acc:sos2=0) no longer limits an unbounded argument to +-1e6 silently: this amplified the feasibility tolerance by up to 1e6 and could give wrong solutions. An unbounded argument is now an error; finite bounds are used as given. New option cvt:pl:domain sets the limit explicitly, with a warning when it applies (use with caution). With native SOS2, unbounded sides of the argument are now modeled exactly by ray variables instead of extending the end segments to 1e6; cvt:pl:domain restores the extension.
  • Fix: AMPLSSolve (AMPLS C API) called the solver once, skipping the iterations of the multiobjective emulator (obj:multi=2, or obj:multi=1 without native support); postsolving the results then read uninitialized objectives and crashed. It now solves as drivers do. This affected ASL drivers run with mp2nl=1.
  • AMPLS C API: AMPLSReportError writes the .sol file with a solve result and message after a failed AMPLSLoadNLModel or AMPLSReportResults, as drivers do on errors.
  • Fix: 64-bit integer options rejected values above 2^31 ("this number is larger than an int"); they are now parsed as 64-bit integers.
  • Fix: the IIS of a model with logical constraints or if-then-else expressions marked many unrelated constraints (upp): every constraint that used a constant, e.g. the implicit else-0 of if ... then, received the IIS status of MP's shared fixed variable for that constant. Constants are no longer linked to the constraints using them. Constraints sharing a subexpression that is in the IIS are all reported.
  • Fix: with several objectives (obj:multi), mip:return_gap and mip:bestbound stopped debug builds with an assertion in ReportSuffix; release builds were unaffected. The objective suffixes .relmipgap, .absmipgap and .bestbound carry the value for the first objective and 0 for the others.
  • Option cvt:pre:boundsbest value 4: best-known bounds for all variables when the submitted model has no nonlinear or quadratic constraints or objectives, hard bounds otherwise. MP2NL sets it in the configurations of the nonlinear subsolvers (knitro, ipopt, conopt, snopt, minos, lgo, bonmin, couenne, baron, lindoglobal): on MILPs the bounds speed up their branch-and-bound, on MINLPs the bounds of auxiliary results degrade their barrier subproblems.
  • Option cvt:pre:boundsbest has values 2 and 3: 2 submits the best-known bounds of all result variables except those of smooth nonlinear expressions (exp, log, pow, trigonometric functions, divisions, products), 3 only those. The decision whether a result variable with a significant best-known bound stays explicit follows the same per-variable rule.
  • Option cvt:pre:auxint (default 0): auxiliary result variables whose integrality is implied by their defining constraint (if-then-else, min/max/abs, count/numberof, integer linear and quadratic functions of integer arguments) are submitted as continuous; MP keeps their integrality for its reformulations. Results of logical constraints, indicator binaries and arguments of logical constraints submitted to the solver stay integer. Solvers without continuous variables (CP-SAT) default to 1. Fixed variables made continuous (cvt:pre:continuous_fixed_vars) no longer change MP's internal variable types.
  • Fix: the big-M linearization of an indicator b==v ==> c'x >= d (or <=) dropped the row when d equalled the body's best-known bound, although that bound is not submitted by default (cvt:pre:boundsbest=0). The solver then saw, e.g., the result of if b then x else 0 without its lower bound 0, and objectives containing it were underestimated. The plain row is now kept.
  • cvt:pre:disj2ind: a variable reused as an indicator binary (an if-then-else result in {0, 1}) gets hard bounds [0, 1] (Xpress: "Non binary variable used as an indicator binary").
  • Option cvt:pre:disj2ind (default 11): convert disjunctions literal || C with a single-use comparison C, which is how AMPL sends implications b==1 ==> C, into indicators b==1 ==> C, without auxiliary binaries. Also for conjunctions C1 && C2 and for x != k; x == k awaiting the unary encoding decision is absorbed only if no encoding is chosen.
  • Option cvt:pre:vbnd (default 3): a comparison x (cmp) c decided by the domain of x reuses an existing binary. Bit 1: variable-bound constraints x >= Lb*, x <= Ub* (b binary; AMPL's encoding of x in {0} union interval[L, U]), with 0 <= c <= L: b or its negation. Bit 2: root disjunctions x <= a || x >= b (e.g., x == 0 || x >= L), a < c < b: the comparison x >= b or its negation.
  • cvt:pre:disj2ind bits 16 (an AND of comparisons in positive context, e.g., a disjunct of a multi-term disjunction, becomes indicators on the AND's result) and 32 (a static disjunction with an AND argument is distributed into clauses); AND members which are not comparisons are implied by clauses.
  • Solution check: in the realistic mode, evaluate recomputed conditional comparisons within sol:chk:feastol; recompute only eliminated auxiliary variables.
  • Emulate mip:basis for linear algebraic MIPs in COPT, MOSEK, HiGHS, and SCIP. Add independent fixed-model limits mip:basis:timelimit and mip:basis:iterlimit, including native fixed-model paths.
  • obj:multi:stats with multi-objective emulation: report .objpass_objbound separately from .objpass_mipgap; report the statistics also when a pass stops the solve (.objpass=0, .objpass_result=-1 and NaN values for the passes not run).
  • Fix multi-objective emulation when objectives of different senses share a priority level: the sign of the weight was flipped twice for objectives of a sense different from the first objective of the level, giving wrong solutions.

20260908

  • Set membership operator discouraged during rework, see Modeling Guide.
  • Don't fail when mip:return_gap or mip:bestbound ask for values but they are not available (warning issued).
  • Option obj:multi:stats=1 outputs multi-objective pass statistics in the .objpass[_...] suffixes.
  • Fix a bug in complementarity reformulation with cvt:compl=2,3.
  • Options mip:plateau:... control MIP tailing-off in supporting solvers. Can be used in multi-objective solves, see option obj:multi:options.
  • Option cvt:nlobj controls whether nonlinear objective terms are passed to the objective, vs moved to a constraint or piecewise-linearly approximated. Setting cvt:nlobj=0 cvt:quadobj=0 enables native multiobjective method (obj:multi=1) if only supported for linear objectives.
  • Option cvt:pre:unnest: bit 4 (value 16) turns on inlining of algebraic subexpressions in indicator constraints. When it is off, the body of indicator constraints is always just a variable (possibly assigned to a larger expression). On by default.

20260624

  • CMake option MP_LINK_WITH_SHARED_LIB links the driver as a shared library (Gurobi, IBM ILOG CPLEX, Cardinal Optimizer, FICO Xpress; for them, it's the default).
  • Option debug=1 outputs all options in problem suffixes option___....
  • Option cvt:pow2_as_qp prefers quadratization of the ^2 operator for affine arguments (default).
  • Faster parsing of quadratic expressions with defined variables.

20260515

  • Fixed a bug in solve result name in the Multi-Objective Emulator.
  • Similar to inlining, nonlinear expressions with nontrivial external bounds are explicified into an auxiliary variable, depending on options cvt:pre:boundsbest and cvt:pre:boundlogarg.
  • Fixed a debugging feature on Windows.
  • Fixed a non-critical bug in the inlining of algebraic subexpressions resulting from redefinitions, option cvt:pre:unnest. In particular with cvt:pre:boundsbest=1, subexpressions with stronger result bounds are not inlined.

20260414

  • Fixed a bug in the multi-objective emulator (see option obj:multi) when nonlinear terms were not reset for consecutive objectives.
  • Option cvt:pre:continuous_fixed_vars to declare fixed variables continuous. Default 1.
  • Option cvt:pre:boundsbest defaults to making auxiliary subexpression result variables free, unless the result is constant. This allows stronger nonlinear presolve.
  • Additional case where the signpow function is recognized: abs(C*x)*x (C constant, x variable).
  • Auxiliary variables and constraints receive names describing the redefinition path. Such names are only given if AMPL option (solver_)auxfiles rc is set, or solver option cvt:names=2,3 provided.
  • Added option tech:stats to return solution statistics via problem suffixes.

20260109

  • Remove excessive '_' symbols in output model constraint and variable names.

20251213

  • Recognize the logistic function 1/(1+e^(-x)), option cvt:pre:logistic.
  • Recognize the signpow function sign(x)*x^k from several kinds of algebra, such as abs(x)*x^k. Option cvt:pre:signpow.
  • iis:find now the main name for option alg:iisfind (which is now a synonym).
  • Option obj:multi:options: fix integer-valued options.

20251121

20251021

  • Option cvt:expr:nlassign adjusting expression inlining. Improves problem building performance for some nonlinear solvers on large models.

20251015

  • Option cvt:pre:ctx2count is deprecated, with a new default value 7. Its former behavior is specialized and generalized by the new option cvt:pre:ctx2bndeq.
  • Option cvt:pre:ctx2bndeq controls whether conditional comparisons with upper/lower bounds are reformulated in exact context, vs. always mixed. Default yes, value 1.
  • Workaround for #270 (redefinition of an expression as an existing expression can lose connection to the original).
  • Fixed multiobj=2 for a single objective. Running multiobjective for a single-objective model makes sense with a negative .objweight suffix.
  • Options cvt:pre:ineq2bndeq, cvt:pre:ineq2related control unification of conditional equlaities and inequalities.
  • cvt:compl:eps=1e-9 new default.
  • alg:start=3 default (new value), which passes both the initial guesses and basis for LP (previously basis only).
  • Option cvt:uenc:negctx:max disregards lb(x), ub(x).
  • Options cvt:pre:ctx:abs etc. control propagation of expression context.

20250814

  • Fixed reformulation of numberof and alldiff. Previously could produce an inefficient reformulation.
  • Option cvt:pre:sort:
    • Sorting and eliminating duplicates in arguments of AND, OR, MIN, MAX. This is necessary for some solvers.
    • Sorting arguments of COUNT, ATLEAST, EXACTLY, ATMOST, NUMBEROF, ALLDIFF.
  • Option cvt:pre:unnest extended for MIN, MAX (bit 3).

20250811

  • The legacy fmt::MemoryWriter has double output precision of 17 significant digits (previously 6).
  • Option cvt:pre:boundlogarg (default 0) to bound arguments of logarithms nonnegative. Previously always done, sometimes deteriorating performance of nonlinear solvers.

20250801

  • Tolerances set by options cvt:pre:feastol, cvt:pre:feastolrel both need to be violated to produce a warning on contradicting variable/constraint bounds. Previously the preprocessor failed on any violation, without letting the solver try.
  • Options cvt:compl, cvt:compl:eps control complementarity reformulations.

20250617

  • Option cvt:unnest: bits 1 and 2 switch on inlining of linear and quadratic subexpressions produced during reformulations (by default on).
  • Options cvt:pre:ctx2ineq, cvt:pre:ctx2count to control context propagation into conditional comparisons #267.

20250616

  • Multi-objective emulator: added support for objective-specific options via objective suffixes beginning with option_.

20250429

  • Fixed a bug in parsing of quadratic expressions, which could wrongly parse products of unequal linear expressions, such as (x-3)*(x-z-5).

20250424

  • Option alg:sens returns synonym suffixes .down/.up/.current for objective coefficients and .down/.up for right-hand sides.
  • Improved parsing speed of large sums of quadratic and polynomial expressions.
    • Setting option cvt:qp2passes=0 switches to a simpler method, usually slightly slower.
    • Option cvt:multoutcard to limit the size of out-multiplied QP expressions. Can improve speed on large models, but prone to numerical issues.

20250308

  • Fixed a bug in monotonicity presolve which could have led to wrong models using a^x with a<1.
  • Reduced memory fragmentation for linear and quadratic expressions.
  • Option tech:outlev_mp=1 (implied by tech:outlev=1) prints initial and transformed model statistics.
  • Improved presolve for disequality: more cases when no logical disjunction is needed.

20250205

  • Fixed a bug in model reformulations which could cause some constraints to be lost, see issue 248.
  • Option alg:sens=1 now returns suffix .sensobj with current objective coefficients in the solver model (this is the model corresponding to AMPL command solexpand.)
  • Option acc:_all is overridden by individual acceptance options, e.g., acc:or.
  • Substitute AMPL defined variables into linear, quadratic, and polynomial expressions (option cvt:dvelim). This can simplify quadratic and polynomial models (linear substitutions are already performed by AMPL, see AMPL options linelim and *substout.)
  • cvt:prod=7 default for LP and convex solvers, logicalizing also products of just 2 binary variables.
  • More presolve for logical expressions (Not, And, Or, Indicator), and new options to control some of them: cvt:pre:ineqresult, cvt:pre:ineqrhs.
  • Fix lower bound calculation of the division result.
  • [BREAKING] Option acc:pow now affects only expressions x^y with both x, y variable; previous meaning of acc:pow is now with acc:powconstexp.
  • Option tech:writemodel:index to choose the iteration when solver model is exported in the multi-objective emulator.
  • SCIP (and any solver with linear objective and non-linear constraints): improve reformulation of QP objectives.
  • Fix reformulation of non-linear objective expressions for the multi-objective case (option obj:multi) when negative objective weights are used (obj:multi:weight).

20240724

  • Option acc:_all
    • Useful to disable all reformulations (acc:_all=2), or force linearization (acc:_all=0).
  • Faster input of quadratic expressions.

20240617

  • Multi-objective emulator
    • obj:multi=2 forces emulation, even if MO natively supported.
      • Fixed a bug in the objective degradation suffixes .objabstol, .objreltol.

20240604

  • Presolve division by constant, resulting in fewer constraints
  • Fix no-solution case in multi-objective emulator

20240529

  • Multi-objective emulator
    • All flat MP solvers support multi-objective mode (obj:multi=1), either natively, or via emulation.
    • Suffixes .objpriority, .objweight, .objabstol, .objreltol.
    • [BREAKING] Default intuitive handling of .objweight, see option obj:multi:weight, even when natively supported.

20240429

  • Options report_times and timing.
    • Options timing and report_times now have the same effect: setting their value to 1 returns basic timing information, setting it to 2 returns more granular info.

20240320

  • SOS constraints.
    • Fixed handling of SOS2 constraints created by AMPL as reformulations of PL expressions (option pl_linearize 1, default; set to 0 to use the solver's native PL functions if supported, or MP linearization.)
    • Disallow repeated weights for SOS constraints (suffixes .sosno/.ref).
  • Reformulation explorer.
    • Upgraded option writegraph exports the reformulation graph which can be explored with the script in support/modelexplore (WIP).
  • Native handling of POW(x, INT).
    • Power expressions with positive integer exponent are passed natively to the solvers accepting them, vs previously quadratic or linear reformulation.
    • For best performance, global solving capability might be needed (e.g., Gurobi: global=1.)
  • Option report_times.
  • Unused acc: options.
    • The constraint acceptance options acc:... for non-handled constraints are ignored (previously triggered error.)
  • NLWPY.
    • Python NL Writer API for MIQP models.
    • Available on PyPI.

20240115

  • Solve result codes.
    • List codes by running (solver) -!
    • [BREAKING] Standardized codes. Major changes:
      • 100-199 (solved?) means solution candidate provided, but can be suboptimal/infeasible
      • 300-349 means unbounded problem but feasible solution returned
      • 400-449 means limit/interrupt but feasible
    • [BREAKING] sol:chk:fail returns code 150 (solved?)
  • Improved translation of SOCP constraints.
    • Options cvt:socp, cvt:socp2qc.
  • Compact solution check warnings
  • Fixed presolve of the power function #226.
  • Fixed graceful exit on Ctrl-C from AMPL in Linux.

20231103

  • Improved translation of logical constraints: inlining of nested disjunctions and conjunctions; fewer auxiliary binary variables.

20231017

  • Fixed a bug in NL reader on Windows.

20230919

  • mp_options. Receive mp_options from AMPL (for all MP solvers). They are parsed before (solvername)_options.
  • Solution checking: relative tolerance sol:chk:feastolrel; options sol:chk:round, sol:chk:prec.

20230831

  • Solution checking, options sol:chk:... (experimental).
  • Preprocess And/Or constraints.

20230817

  • Alternative solutions: solve status equal to that of the final solution.
  • Fixed a bug causing repeated names for auxiliary variables and constraints.
  • Option values can be assigned without '='.
  • Fixed a bug where equivalent conditional comparisons were not unified.

20230728

  • Option 'tech:writesolution' #218.
  • Option 'writeprob' ('tech:writemodel') ASL-compatible.
  • Hint when 'writeprob' fails: use 'writesol'.

20230726

  • Fixed inequalities of integer expressions with non-integer constants, see test_int_non_int.mod.
  • Backend std feature WRITE_SOLUTION.
  • Fixed parsing quoted string options.

20230724

  • Option [solver_]auxfiles rc; transfers names of variables and constraints into the model; (solver)_options 'cvt:names=0-3' controls names.

20230714

  • Print warnings in non-verbose mode too.
  • 'barrier' equivalent to 'barrier=1' for integer options.

20230621

  • Fix quadratic objective with repeated subexpressions.

20230616

  • Smaller reformulations for conditional comparisons.
  • Option cvt:names sets whether to read AMPL variable names or to provide generic names.

20230531

  • Cones: recognize (affine_expr) >= y * exp(z/y) as exponential cone.
  • Cones: recognize xy >= 1 as rotated SOC.
  • Wrong solver options are gracefully reported via solve_message.

20230515

  • Recognize exponential conic constraints. Exponential cones are recognized and passed to the solver, if supported.

20230424

  • Pass variable names if read from a col file with the same name of the nl file being read.
  • Fixed #203: starting solution is now not passed to the solver if empty.

20230321

  • Recognize second-order cones Recognize SOCP constraints from algebra and pass them natively, or transform to quadratics.

20230207

  • Handle boolean constants in ProblemFlattener.

20221228

  • More to #163: ignore SOS with zero weights.

20221222

  • Fixed #195: case-insensitive option synonyms.

  • Fixed #194 Report correct objno for feasibility problems in .sol file, so that AMPL can print "Objective = find feasible solution".

20221211

  • ==> else Implemented implication with 'else': constr1 ==> constr2 [else constr3]

  • PLApproxRelTol, PLApproxDomain Parameters to control piecewise-linear approximation. cvt:plapprox:reltol default value changed from 1e-5 to 0.01.

20221012

  • Piecewise-linear approximation of quadratics Automatic for linear solvers. For convex QP solvers, set the following options: cvt:quadobj=0 cvt:quadcon=0 to linearize nonconvex objective(s) and/or constraints. Recognizing x^2 for stronger univariate approximation.

20220928

  • Piecewise-linear approximation of univariate nonlinear functions Approximation of exp, a^x, x^a, log, log10, trigonometric and hyperbolic functions.

  • Default value of big-M For linearization of logical constraints on variables without finite bounds, option cvt:mip:bigM can provide a default big-M bound.

20220725

  • Solution file export On Windows now creates files with LF only to avoid issues when exporting suffixes to AMPL.

    Multiple solutions export file format amended.

20220720

  • Propagating suffixes via expression trees into flat constraints Partially implemented #184. x-gurobi accepts options 'funcpieces...' and corresponding suffixes which are passed into GRBaddgenconstrExp etc.

    Subexpressions: note that if a subexpression is contained in several constraints, for contradicting suffix values the maximum is taken.

  • Option 'cvt:writegraph' Exporting the flattening / conversion graph in JSON Lines format (WIP).

20220617

  • PowConstraint is reduced to quadratics in some cases For constant non-negative integer exponent and base variables with negative lower bound, PowConstraint is reduced to quadratics (possibly with auxiliary variables). Reason: Gurobi's GRBaddgenconstrPow not accepting negative bases.

  • Context for algebraic constraints Context is now propagated for algebraic constraints. For example, 3x + max(y, z) <= 6 will result in 3 linear constraints. (Earlier this was done for logical constraints and objectives).

20220526

  • Special ordered sets Fixed: SOS are now recognized even if the suffix '.ref' value is integer

20220511

  • Complementarity constraints: also quadratics Complementarity constraints now handle quadratics.

  • Branch develop is used for new code The active development branch is now develop.

  • Convert quadratic range constraints to QuadCon(LE/EQ/GE) Gurobi and COPT do not support quadratic range constraints. (Gurobi's linear ranges are not feature-complete). Conversion of linear range constraints into one-side rhs constraints has been generalized for any algebraic ones.

20220408

  • Complementarity constraints: 1st go Conversion to MIP of complementarity constraints (no quadratics but functional subexpressions ok).

  • DivConstraint and DivConverter_MIP ModelFlattener now receives the Div expression and MIPFlatConverter handles it via quadratics.

  • Fix NL input variable order

  • Reduce default strict comparison tolerance Change cvt:mip:eps default value to 1e-4.

  • Build on MacOS 12.3, in particular on Apple M1 Fixed linking on MacOS 12.3 and FindCPLEX.cmake. For Apple M1, manually set -DCMAKE_OSX_ARCHITECTURES="x86_64" in CMake when building with CPLEX 22.1 because it contains only Intel libraries.

  • Expression maps FlatConverter eliminates subexpressions of all types. A subexpression means here a duplicate expression, such as abs(x+2) occurring several times in the model (here x+2 is a nested subexpression).

  • AMPLS C API C API allowing access to underlying solver API. Replaces the previous Solver C API (solver-c.cc). Toy driver gurobi_ampls exemplifies API usage.

20220216

  • Improved warnings (#161, #163): In verbose mode, FlatConverter / Backend print warning summary before and after solving

  • Fixed solver option parsing in Windows (#160)

  • Reworked Backend / ModelAPI class hierarchy (#162): In particular, also generalized old MP hierarchy (Solver / ProblemBuilder / AppSolutionHandler / SolverNLHandler)

  • Allowing SOS constraints with repeated weights (#163): Although Gurobi states SOS weights should be unique, it accepts them repeated. This happens when AMPL linearizes a PL function with redundant (repeated) slopes. It seems better to use PL functions natively (option pl_linearize 0;).