{
  "id": "mixed-integer-programming",
  "title": "Mixed-integer programming",
  "lang": "en",
  "source_lang": "en",
  "summary": "Mixed-integer programming optimizes decisions involving both discrete choices and continuous quantities under mathematical constraints.",
  "category": "mathematics",
  "subcategory": "applied",
  "type": "concept",
  "aliases": [
    "Mixed-integer optimization",
    "MIP"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/mixed-integer-programming/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/mixed-integer-programming/",
      "json": "https://aiwiki.page/zh/mixed-integer-programming.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Mixed-integer programming is a branch of mathematical optimization in which some decision variables must take integer values while others may vary continuously. It provides a common language for decisions that combine discrete choices—such as opening a facility, selecting a route, or starting a machine—with quantities such as production levels, shipment volumes, or operating times. The word “programming” refers to mathematical planning rather than to computer programming.\n\nThe most widely used form is mixed-integer linear programming, in which the objective and constraints are linear. Mixed-integer quadratic, conic, and nonlinear programming extend the framework to other mathematical relationships. These models are central to operations research because many operational and strategic decisions cannot be represented adequately by continuous optimization alone.\n\nMixed-integer optimization combines modeling with algorithms that search discrete alternatives while calculating bounds on the best achievable objective value. Its practical effectiveness depends on the interaction among formulation strength, continuous relaxations, search, mathematical inference, and numerical computation. A solution method may therefore produce both an implementable decision and a quantitative certificate of how close that decision is to optimality."
    },
    {
      "heading": "Mathematical formulation and scope",
      "text": "A standard mixed-integer linear program, or MILP, has the form\n\n\\[\\begin{aligned} \\min_{x,y}\\quad & c^\\mathsf{T}x+d^\\mathsf{T}y\\\\ \\text{subject to}\\quad & Ax+By\\le b,\\\\ & x\\in\\mathbb{R}^{p},\\\\ & y\\in\\mathbb{Z}^{q}. \\end{aligned}\\]\n\nHere \\(x\\) contains continuous variables, \\(y\\) contains integer variables, and the matrices and vectors specify the constraints and objective coefficients. Equalities, lower bounds, and upper bounds can be included directly or represented through inequalities. Binary variables, restricted to \\(\\{0,1\\}\\), are especially common because they encode yes-or-no decisions. The feasible set consists of all assignments satisfying both the algebraic constraints and the variable-domain restrictions.\n\nStrictly, a mixed-integer model contains both continuous and integer variables. In software and much of the literature, however, “MIP” also names a broader problem class that includes models whose variables are all integer or binary. Integer programming is therefore both a closely related field and, under this broader usage, a special case. Removing every integrality requirement from an MILP produces a linear program. By contrast, adding integer restrictions to a nonlinear model produces a mixed-integer nonlinear program, whose difficulty depends strongly on whether its continuous structure is convex.\n\nA simple fixed-charge production model illustrates the distinction between discrete and continuous decisions. Let \\(x\\ge0\\) be the quantity produced and let \\(y\\in\\{0,1\\}\\) indicate whether production is activated. With capacity \\(U\\), fixed activation cost \\(F\\), and unit production cost \\(c\\), the model may contain\n\n\\[0\\le x\\le Uy, \\qquad \\text{cost}=cx+Fy.\\]\n\nWhen \\(y=0\\), the linking constraint forces \\(x=0\\). When \\(y=1\\), production may range up to capacity. A demand constraint such as \\(x\\ge D\\) completes a basic production decision. The binary variable represents a genuine discontinuity: producing a positive amount triggers a fixed cost that cannot be represented by an ordinary linear cost in \\(x\\) alone.\n\nMixed-integer programming describes a mathematical structure, not a particular interpretation of uncertainty or time. A model may be deterministic, incorporate multiple future scenarios, or represent many planning periods. Its objective function may measure cost, profit, delay, emissions, or a weighted combination. Multiple objectives can also be treated lexicographically or through constraints on secondary criteria. These choices determine the planning problem; integrality determines how discrete decisions enter its mathematical representation."
    },
    {
      "heading": "Relaxations, geometry, and computational complexity",
      "text": "The principal starting point for solving an MILP is its linear programming relaxation, obtained by allowing the integer variables to take real values. For a minimization problem, this enlarges the feasible set and therefore gives a lower bound:\n\n\\[z_{\\mathrm{LP}}\\le z_{\\mathrm{MIP}}.\\]\n\nAny feasible mixed-integer solution gives an upper bound. The difference between these bounds measures the remaining uncertainty about the optimal objective value. A good relaxation does more than approximate the answer: it allows entire families of discrete alternatives to be ruled out without explicit enumeration.\n\nGeometrically, the linear relaxation is a polyhedron, whereas the mixed-integer feasible set retains only those points whose designated coordinates are integral. The convex hull of the mixed-integer feasible points is the ideal continuous representation for optimizing a linear objective. For a nonempty bounded feasible region described by rational linear data, this hull is a rational polytope. An exact description may nevertheless require an enormous number of inequalities, and finding useful portions of that description is a central task in integer optimization.\n\nThe difference between an integer optimum and a relaxation optimum is commonly called an integrality gap. The term may denote an additive difference or a ratio, depending on the problem and sign conventions. For example, consider binary variables satisfying\n\n\\[2y_1+2y_2\\le3\\]\n\nand an objective that maximizes \\(y_1+y_2\\). The integer optimum is \\(1\\), but the continuous relaxation permits \\(y_1=1\\), \\(y_2=\\tfrac12\\), with objective \\(1.5\\). Adding \\(y_1+y_2\\le1\\) closes the gap. Some structured linear systems need no such strengthening: with suitable integral right-hand sides, a totally unimodular constraint matrix guarantees integral vertices. This explains why important network-flow and assignment problems can be solved through linear programming despite their discrete interpretation.\n\nGeneral MILP is NP-hard, and the bounded binary linear feasibility problem is NP-complete. These statements concern worst-case complexity; they do not imply that every large instance is difficult or that every small instance is easy. Lenstra’s 1983 result established that integer linear programming is polynomial-time solvable when the dimension is fixed, with dependence on that dimension outside the polynomial in input length. This contrasts with the general case, where the number of integer variables is part of the input. Practical difficulty also depends on symmetry, numerical conditioning, relaxation strength, and the way constraints interact, so variable counts alone are poor predictors of solution time."
    },
    {
      "heading": "Modeling discrete choices and logical relationships",
      "text": "Binary variables translate many logical statements into linear inequalities. If \\(a,b\\in\\{0,1\\}\\), the implication “\\(a\\) implies \\(b\\)” is represented by \\(a\\le b\\). Mutual exclusion is \\(a+b\\le1\\), while exactly one choice among \\(n\\) alternatives is expressed as\n\n\\[\\sum_{i=1}^{n}y_i=1.\\]\n\nFor a binary variable \\(z\\) representing the conjunction of binary variables \\(a\\) and \\(b\\), the inequalities \\(z\\le a\\), \\(z\\le b\\), and \\(z\\ge a+b-1\\) give an exact formulation. Such constructions connect mixed-integer models with propositional logic, but their computational quality depends on more than logical correctness.\n\nA widely used device is the big-M formulation. To impose \\(a^\\mathsf{T}x\\le b\\) only when \\(y=1\\), one may write\n\n\\[a^\\mathsf{T}x\\le b+M(1-y).\\]\n\nThe constant \\(M\\) must be large enough that the inequality does not remove intended feasible solutions when \\(y=0\\). It should also be as small as the variable bounds and other constraints permit. An unnecessarily large value weakens the linear relaxation and may amplify numerical difficulties. Many solvers provide indicator constraints that express the implication directly, although their internal treatment still relies on bounds, branching, or suitable reformulations.\n\nSome nonlinear expressions admit exact linear representations when a factor is binary. If \\(z=xy\\), \\(y\\in\\{0,1\\}\\), and \\(L\\le x\\le U\\), the four inequalities\n\n\\[\\begin{aligned} z&\\ge Ly,& z&\\le Uy,\\\\ z&\\ge x-U(1-y),& z&\\le x-L(1-y) \\end{aligned}\\]\n\nenforce the product exactly. Piecewise-linear costs and physical relationships can likewise be modeled using segment-selection binaries or special ordered sets. These representations are useful for volume discounts, tariff schedules, production modes, and approximations to smooth functions, although an approximation remains distinct from an exact representation of the original relationship.\n\nAlternative formulations can describe the same discrete decisions while yielding very different relaxations. In facility location, for example, a single aggregate capacity constraint may allow fractional facility openings to support implausibly large individual customer assignments; additional assignment-to-opening inequalities can strengthen the relaxation. Formulations may trade more variables and constraints for better bounds. Disaggregated models preserve distinctions that aggregate models hide, while extended formulations introduce auxiliary variables to describe a useful projection of a higher-dimensional polyhedron. Another important choice is whether to eliminate interchangeable labels. Identical machines or vehicles can produce many equivalent solutions, and valid symmetry-breaking constraints can prevent the algorithm from repeatedly exploring these alternatives."
    },
    {
      "heading": "Branch-and-bound and the search for optimality",
      "text": "Branch-and-bound organizes the discrete search into a tree of subproblems. At each node, the algorithm solves a relaxation of the original problem augmented by the restrictions defining that node. Suppose a relaxation returns \\(y_j=3.4\\) for a variable that must be integer. Branching creates two children:\n\n\\[y_j\\le3 \\qquad\\text{and}\\qquad y_j\\ge4.\\]\n\nEvery integer-feasible point in the parent belongs to one of these children, while the current fractional point belongs to neither. Repeating this process partitions the remaining search space without losing any feasible integer solution.\n\nFor minimization, a node can be discarded if its relaxation is infeasible, if its lower bound is no better than the objective of the best known feasible solution, or if an optimal relaxation solution already satisfies all integrality restrictions. The best known feasible solution is called the incumbent. Across all active nodes, the smallest valid lower bound provides a global lower bound \\(L\\), while the incumbent provides an upper bound \\(U\\). If the search closes the gap, optimality is established to the relevant numerical tolerances. If a time limit interrupts the search, the pair \\((L,U)\\) still describes what has been proved.\n\nBranching and node-selection rules strongly affect performance. Strong branching tentatively evaluates candidate branches to estimate their effect on the relaxation bound. Pseudocost branching learns from previous branches how much improvement tends to follow changes in a variable’s bounds. Reliability branching combines initial probing with accumulated estimates. Best-bound node selection emphasizes improving the global certificate, whereas depth-first exploration can use less memory and may find complete integer solutions sooner. Modern methods balance these goals rather than applying one rule uniformly throughout the search.\n\nFeasible solutions are often found by dedicated heuristics rather than by reaching integral relaxation optima. Simple rounding is unreliable because it can violate capacities, conservation equations, or logical relationships. More sophisticated procedures repair rounded points, search neighborhoods around an incumbent, or fix variables on which an incumbent and a relaxation agree. The feasibility pump alternates between a continuous feasible point and an integer point in an attempt to reconcile feasibility and integrality. These methods can rapidly improve the upper bound, but they do not independently prove optimality. On bounded integer domains, exhaustive branching provides a finite conceptual search; practical success comes from avoiding almost all of that enumeration through bounds and inference."
    },
    {
      "heading": "Cutting planes and polyhedral strengthening",
      "text": "A cutting plane is an inequality valid for every feasible mixed-integer solution but violated by a point in the current relaxation. Adding it removes an inadmissible fractional solution without eliminating legitimate discrete decisions. Cuts can tighten the root relaxation before branching begins, and they can also strengthen relaxations at nodes deeper in the search tree. The combination of cuts with branch-and-bound is known as branch-and-cut and is the dominant general framework for solving MILPs.\n\nA simple example comes from a binary knapsack constraint,\n\n\\[\\sum_i a_i y_i\\le b,\\qquad y_i\\in\\{0,1\\},\\]\n\nwith nonnegative weights. If a subset \\(C\\) satisfies \\(\\sum_{i\\in C}a_i>b\\), not every item in \\(C\\) can be selected. The resulting cover inequality is\n\n\\[\\sum_{i\\in C}y_i\\le |C|-1.\\]\n\nFor weights \\(6,6,6\\) and capacity \\(10\\), each pair forms a cover, yielding \\(y_i+y_j\\le1\\). A further valid inequality is \\(y_1+y_2+y_3\\le1\\), which is stronger than the pairwise inequalities alone. Such examples show why the strength and combination of inequalities matter, rather than simply their number.\n\nGeneral-purpose cut families include Gomory mixed-integer cuts, mixed-integer rounding cuts, and split cuts. They exploit the incompatibility between continuous relaxation points and integrality restrictions. Problem-specific cuts use additional structure. In the traveling salesman problem, degree equations alone allow several disconnected cycles. Subtour-elimination inequalities prevent these disconnected tours, and stronger families can further improve the relaxation. A separation procedure searches for a violated inequality in a family; it need not list every possible inequality in advance. Efficient separation allows a compact working model even when the complete mathematical formulation contains exponentially many constraints.\n\nNot all valid cuts are computationally helpful. Each added inequality enlarges the continuous subproblem, and cuts with poorly scaled coefficients can damage numerical reliability. Solvers therefore assess measures such as violation, sparsity, and similarity to existing cuts, while removing cuts that cease to be useful. A globally valid cut applies throughout the tree; a locally valid cut may depend on bounds imposed at a particular node and is valid only within the corresponding subtree. This distinction is essential for correctness. Cutting-plane theory seeks strong descriptions of integer-feasible sets, while solver engineering must balance bound improvement against the cost of maintaining and solving increasingly large relaxations."
    },
    {
      "heading": "Decomposition and large structured models",
      "text": "Large mixed-integer models often contain linking decisions that couple many otherwise independent subproblems. Benders decomposition, introduced by Jacques F. Benders in 1962, exploits a partition between master variables and subproblem variables. In a common application, the master chooses discrete infrastructure decisions, while continuous subproblems determine operating costs. For\n\n\\[\\min_y\\{f^\\mathsf{T}y+Q(y)\\}, \\qquad Q(y)=\\min_x\\{c^\\mathsf{T}x:Wx\\ge h-Ty,\\ x\\ge0\\},\\]\n\nthe subproblem’s dual solutions generate inequalities that approximate the recourse function \\(Q(y)\\) in the master problem.\n\nAn optimal dual solution produces an optimality cut linking a master variable representing operating cost to the discrete decisions \\(y\\). If a proposed \\(y\\) makes the subproblem infeasible, an appropriate dual infeasibility certificate can produce a feasibility cut. The process alternates between proposing master decisions and correcting their estimated consequences. Its effectiveness depends on generating informative cuts, maintaining valid bounds, and controlling repeated solutions of similar subproblems. Logic-based Benders decomposition extends the underlying idea beyond linear subproblems by deriving valid explanations of subproblem outcomes.\n\nDantzig–Wolfe decomposition, developed by George Dantzig and Philip Wolfe in 1960, instead represents blocks of decisions through feasible patterns or extreme points. A master problem chooses combinations of these patterns, while pricing subproblems identify new patterns with improving reduced cost. This is column generation. In cutting-stock models, a column may describe a feasible pattern for cutting smaller pieces from a stock roll. In vehicle routing, a column may represent an entire feasible route. The full collection of columns can be enormous, so generating only useful ones avoids building the entire formulation.\n\nSolving the continuous master by column generation does not generally enforce integer decisions. Combining column generation with branching gives branch-and-price, and adding cuts yields branch-price-and-cut. Branching decisions must be chosen so that pricing remains tractable. A related approach, Lagrangian relaxation, moves selected constraints into the objective with multipliers and may separate the model into easier subproblems. These methods are particularly important in multiscenario planning, where many scenario-specific operational models share common first-stage decisions. Their advantages are structural rather than automatic: decomposition can reduce memory requirements and exploit parallelism, but weak cuts, difficult pricing, or slow coordination can offset those gains."
    },
    {
      "heading": "Quadratic, conic, and nonlinear extensions",
      "text": "Mixed-integer quadratic programming allows a quadratic objective, typically\n\n\\[\\min\\ \\tfrac12 z^\\mathsf{T}Qz+c^\\mathsf{T}z,\\]\n\nsubject to linear constraints and integrality restrictions on selected components of \\(z\\). When \\(Q\\) is positive semidefinite, the continuous relaxation is a convex optimization problem. Such models arise when discrete selections interact with quadratic penalties, including tracking error, smoothness, and variance. Quadratic constraints lead to mixed-integer quadratically constrained programming, with convexity again determining whether continuous subproblems can be solved globally by convex methods.\n\nMixed-integer conic models combine discrete variables with constraints such as\n\n\\[\\|Az+b\\|_2\\le c^\\mathsf{T}z+d.\\]\n\nSecond-order cone constraints can express certain norm bounds, robust optimization counterparts, and relationships involving quadratic quantities. Their convex continuous relaxations provide valid global bounds within branch-and-bound. The important distinction is not simply whether a model contains nonlinear notation, but whether its relaxation has a structure for which reliable global optimization is available.\n\nMixed-integer nonlinear programming, or MINLP, permits more general nonlinear objectives and constraints. For convex MINLP, outer approximation combines nonlinear subproblems at fixed integer assignments with a mixed-integer linear master. If \\(g\\) is differentiable and convex, its first-order approximation satisfies\n\n\\[g(x)\\ge g(\\bar x)+\\nabla g(\\bar x)^\\mathsf{T}(x-\\bar x).\\]\n\nConsequently, the linearized inequality\n\n\\[g(\\bar x)+\\nabla g(\\bar x)^\\mathsf{T}(x-\\bar x)\\le0\\]\n\nis necessary for \\(g(x)\\le0\\) and can contribute to an outer relaxation. Under suitable assumptions, repeated linearizations and bounds support a finite solution procedure. Other approaches solve convex nonlinear relaxations directly within a branching tree.\n\nNonconvex MINLP requires additional care because a local continuous optimum does not provide a valid global lower bound for minimization. Global methods construct convex underestimators and may branch on continuous variables as well as integer variables. McCormick envelopes provide standard convex relaxations for bounded bilinear terms; these become tighter as variable intervals shrink. Perspective reformulations can strengthen convex on/off costs by exploiting the connection between an activation variable and its associated continuous activity. These extensions preserve the central mixed-integer principle—combine discrete search with valid bounds—but the continuous bounding problem can become substantially more difficult than an LP. A solver’s claim of local feasibility must therefore be distinguished from a certificate of global optimality."
    },
    {
      "heading": "Numerical computation, certificates, and software",
      "text": "Practical solvers begin with presolve, which simplifies a model while preserving its relevant mathematical properties. Presolve can fix variables, tighten bounds, remove redundant constraints, substitute equalities, and detect contradictions. Integrality itself strengthens propagation: if an integer variable satisfies \\(y\\le3.8\\), it also satisfies \\(y\\le3\\). Repeated propagation across constraints can reveal implications that were not apparent in any single row. Probing temporarily assigns a binary variable and follows the resulting deductions, potentially discovering fixings, implications, or infeasible assignments.\n\nMost large-scale computation uses floating-point arithmetic, so feasibility is tested within tolerances rather than through exact symbolic equality. An apparent integer value may differ slightly from its nearest integer, and constraint residuals may be accepted within a specified threshold. Large coefficient ranges and oversized big-\\(M\\) values can magnify the practical effects of these tolerances. In a constraint \\(x\\le My\\), a tiny nonzero value of \\(y\\) can permit a noticeable value of \\(x\\) when \\(M\\) is very large. Scaling, valid tight bounds, and careful formulation therefore affect both speed and the reliability of computed solutions.\n\nSolver reports distinguish several outcomes. A feasible incumbent need not be optimal; a time limit is not a proof of infeasibility; and absence of an incumbent does not by itself establish that no feasible solution exists. For minimization, a reported lower bound \\(L\\) and incumbent value \\(U\\) imply an absolute objective gap \\(U-L\\). Relative-gap formulas vary among solvers, especially near zero or when objective values have different signs. Exact arithmetic and independently checkable proof certificates can strengthen verification for suitable problem classes and solver workflows, but such certification is distinct from ordinary floating-point optimization.\n\nCommercial systems include CPLEX, Gurobi, and FICO Xpress; open-source systems include SCIP, CBC, and HiGHS. Modeling environments such as AMPL, GAMS, Pyomo, and JuMP separate mathematical model construction from the underlying solution engine. Features such as callbacks, warm starts, and parallel search permit substantial customization. For infeasible models, conflict analysis or irreducible infeasible subsystem tools can identify subsets of incompatible requirements, although the smallest explanatory subset is not necessarily found. Reproducible performance comparisons require model files, software versions, hardware details, settings, stopping criteria, and attention to variability from randomization or parallel execution."
    },
    {
      "heading": "Applications and the interpretation of solutions",
      "text": "Facility location is a canonical mixed-integer application. Binary variables \\(y_j\\) indicate whether facilities are opened, while continuous variables \\(x_{ij}\\) specify shipments from facility \\(j\\) to customer \\(i\\). A basic formulation is\n\n\\[\\begin{aligned} \\min\\quad& \\sum_j F_jy_j+\\sum_{i,j}c_{ij}x_{ij}\\\\ \\text{subject to}\\quad& \\sum_jx_{ij}=d_i &&\\text{for each customer }i,\\\\ &\\sum_ix_{ij}\\le K_jy_j &&\\text{for each facility }j,\\\\ &x_{ij}\\ge0,\\quad y_j\\in\\{0,1\\}. \\end{aligned}\\]\n\nThe model balances opening costs against transport costs and capacity. Extensions add service-distance limits, facility types, multiple commodities, or uncertain demand. An apparently modest change—such as requiring each customer to use only one facility—introduces additional discrete assignment decisions and can substantially change the formulation.\n\nProduction planning and scheduling combine setup, sequencing, and quantity decisions. Lot-sizing models charge a setup cost when production occurs in a period, while inventory-balance equations link production and demand over time. Scheduling models can represent ordering by binary precedence variables or assign jobs to discrete time slots. Precedence formulations may be compact but depend on strong bounds for disjunctive timing constraints. Time-indexed formulations can give strong relaxations but grow with the planning horizon and time resolution. The choice of representation therefore affects both the meaning of the schedule and the computational burden.\n\nIn electricity-system unit commitment, binary variables represent whether generating units are operating, starting, or shutting down, while continuous variables represent output and reserves. Constraints encode ramp limits, minimum up and down times, demand balance, and transmission approximations. In transportation, mixed-integer models select routes and assign vehicles or crews while coordinating times and capacities. Vehicle-routing models often combine binary arc decisions with continuous load or arrival-time variables. Network design similarly combines discrete investment choices with continuous flows. These applications illustrate why purely continuous plans may be operationally meaningless even when they satisfy aggregate resource equations.\n\nMixed-integer methods also appear in statistical learning and automated decision systems. A cardinality constraint can limit the number of selected predictors, while binary variables may encode classification rules or tree structure. Here the optimization certificate concerns the specified empirical objective, not predictive validity or causal interpretation. More generally, mathematical optimality is conditional on the model: it does not establish that demand forecasts are correct, omitted constraints are irrelevant, or chosen objective weights capture every institutional concern. Sensitivity analysis is also less straightforward than in linear programming because small data changes can switch discrete configurations. Scenario comparisons, repeated optimization, and evaluation of alternative feasible decisions are therefore important parts of interpreting mixed-integer results."
    },
    {
      "heading": "Historical development and research directions",
      "text": "Mixed-integer programming developed from the interaction of economic planning, combinatorial optimization, and electronic computing. George Dantzig’s simplex method, introduced in 1947, made large-scale linear programming practicable and supplied the continuous optimization foundation later used for integer bounds. In 1954, Dantzig, Ray Fulkerson, and Selmer Johnson published a solution of a 49-city traveling salesman instance using linear programming and additional inequalities. That work demonstrated the value of exploiting a relaxation while progressively excluding structures incompatible with the desired discrete solution.\n\nRalph Gomory introduced general cutting-plane methods for integer programming in the late 1950s. Ailsa Land and Alison Doig’s 1960 paper presented a systematic discrete programming procedure that became a foundation of branch-and-bound. The Dantzig–Wolfe decomposition paper also appeared in 1960, followed by Benders decomposition in 1962. These developments established several complementary principles: represent discrete structure through inequalities, partition alternatives through branching, and exploit separable structure through decomposition. None alone solved the general practical problem, but their combination created the basis for modern methods.\n\nDuring subsequent decades, research moved increasingly toward integrating these ideas into solver systems. Polyhedral studies identified strong inequalities for routing, matching, facility location, and other structured problems. Improved LP algorithms, sparse linear algebra, presolve, branching rules, primal heuristics, and cut management collectively transformed computational performance. The MIPLIB benchmark library, first released in 1992 and updated in later collections, provided shared test instances for comparing methods. Benchmarking also made clear that progress must be assessed across varied structures rather than through a single favorable example.\n\nCurrent research continues to examine the interaction between mathematical structure and adaptive computation. Topics include stronger formulations, automated decomposition, symmetry handling, proof logging, parallel search, and methods for models that change repeatedly over time. Machine learning has been applied to branching, cut selection, and other solver decisions, but learned policies must still operate within a framework that preserves valid bounds and correctness. Another persistent issue is explaining performance: formulations of similar size may behave very differently, and changes that improve one instance family may harm another. These questions link the theory of discrete optimization to numerical analysis, software architecture, and the empirical study of algorithms."
    }
  ],
  "links": [
    "mathematical-optimization",
    "operations-research",
    "feasible-set",
    "integer-programming",
    "linear-programming",
    "objective-function",
    "polyhedron",
    "convex-hull",
    "np-hardness",
    "propositional-logic",
    "branch-and-bound",
    "cutting-plane-method",
    "traveling-salesman-problem",
    "lagrangian-duality",
    "convex-optimization",
    "floating-point-arithmetic",
    "sensitivity-analysis",
    "machine-learning"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}