Operations research is the mathematics of designing and improving complex decisions under constraints. It combines optimisation, probability, statistics, graph theory, simulation and decision analysis to help allocate scarce resources, schedule activities, route flows, manage queues and compare competing policies.
The field begins when a system contains many interacting choices. One delivery may be easy to plan. Hundreds of deliveries sharing vehicles, roads, drivers, time windows and capacity become a system problem. One classroom timetable may be manageable by inspection. An institution with rooms, teachers, subjects and conflicts needs a more disciplined architecture.
Operations research does not promise that one equation can capture every human value. Its strength is narrower and more useful: make the decision structure explicit, quantify important constraints and consequences, find or compare feasible plans, test sensitivity and state clearly what the model has and has not proved.
Series route: Mathematics Learning Hub → How Mathematics Works → Operations Research. Core mathematical routes include Mathematical Optimisation, Graph Theory, Probability and Stochastic Processes.
1. What operations research actually does
An operations-research study translates a real decision into a mathematical or computational model, solves or explores that model, checks whether the result is robust enough to matter, and returns the result to the operating system that created the question.
Typical questions include: Which jobs should be assigned to which machines? How should vehicles be routed? How much inventory should be held? How many servers are needed to keep queues acceptable? Which projects should receive limited capital? Which timetable minimises conflicts while preserving required coverage?
The common pattern is not one formula. It is structured choice under scarcity and interaction.
2. The first model is a map, not the territory
A model deliberately omits detail. The goal is to retain the detail that changes the decision while discarding detail that adds complexity without changing the relevant conclusion.
This makes modelling an act of compression. A road network becomes vertices, edges, capacities and travel times. A workforce becomes skills, shifts, availability and task requirements. Inventory becomes demand, lead time, holding cost and shortage consequences.
The danger is false compression: remove something that actually controls feasibility or value, and the optimiser can return a beautiful answer to an incomplete world.
3. Decisions, parameters and states must be separated
Decision variables are quantities the planner can choose. Parameters describe the current scenario. States describe the condition of a dynamic system and may change as actions occur.
For a delivery model, the route is a decision, travel time estimates are parameters, and current vehicle location may be a state. Mixing these categories makes it difficult to know what the model is actually controlling.
4. Linear programming is the foundational allocation model
A linear programme has a linear objective and linear constraints. It can represent production, blending, resource allocation, transportation and many planning tasks where contributions add approximately and interactions are captured by linear balances.
In standard form, the feasible set is a convex polyhedron. This geometry gives linear programming powerful structure: local improvement can be organised globally, and optimal solutions can be certified through duality.
5. A simple allocation example
Suppose two activities use limited time and preparation resources. Let x and y be hours assigned to them. A planning score may be maximised subject to capacity and minimum-coverage constraints.
Maximise 3x + 2y subject to x + y ≤ 6 2x + y ≤ 8 x ≥ 1, y ≥ 1
The allocation x=2, y=4 is feasible and scores 14. A certificate follows because 3x+2y=(x+y)+(2x+y)≤6+8=14. This is the same optimisation structure used in the companion Mathematical Optimisation article, now viewed as one operations-research building block.
6. Duality gives resources shadow values
The dual problem attaches values to constraints and asks whether those values can bound the original objective.
A shadow price can estimate how much the optimal objective would improve from a small increase in a binding resource, within the range where the same local structure remains valid.
This makes operations research useful not only for “what should we do?” but also “which scarce resource is worth expanding next?”
7. Integer programming handles indivisible choices
Vehicles, people, machines, rooms and projects often cannot be divided fractionally. Integer and binary variables enforce those realities.
Binary variables are especially powerful because they can encode logic: open or closed, assigned or not assigned, selected or rejected.
The price is computational difficulty. A continuous relaxation may be easy while the integer version creates an enormous combinatorial search space.
8. Branch-and-bound combines search with mathematical bounds
Branch-and-bound divides an integer problem into subproblems while using relaxations to prove that some branches cannot beat the best known feasible solution.
It therefore avoids testing every combination blindly. Search is guided by certificates.
The distinction between feasible incumbent and proven optimum remains important. A solver may have an excellent plan before it has closed the optimality gap.
9. Network models turn flow into structure
Roads, pipelines, supply chains and communication systems can be represented by graphs with capacities, costs and directions.
Shortest-path models minimise route cost. Maximum-flow models maximise throughput. Minimum-cost flow combines conservation of flow with costs and capacities.
The Graph Theory branch provides the structural language; operations research attaches decisions and objectives.
10. Transportation models allocate supply to demand
A transportation model decides how much to send from supply nodes to demand nodes while minimising cost or another objective.
Each source has capacity, each destination has requirement, and each route has cost. The model preserves balance across the network.
Adding fixed vehicle costs, time windows or route sequencing can transform the problem into a harder integer or vehicle-routing model.
11. Assignment problems match entities one-to-one
Assignment models match workers to jobs, teachers to classes, machines to tasks or crews to routes.
Each match can carry cost, time or preference. Constraints prevent one entity from being assigned incompatibly.
The bipartite graph view makes the structure visible; linear and combinatorial algorithms exploit it efficiently.
12. Scheduling adds time and precedence
Scheduling asks when tasks should occur and which resources should perform them.
Tasks may require prerequisites, deadlines, setup times and shared machines. These interactions create disjunctive constraints: if two tasks need the same machine, one must occur before the other.
Good schedules therefore coordinate both allocation and sequence.
13. Critical-path methods expose project bottlenecks
A project network can represent activities and precedence relationships. The critical path is a longest-duration chain whose delay delays the project under the model.
Activities with slack may move without changing the completion time, while critical activities have no such reserve.
This provides more information than one predicted completion date: it identifies where delay matters most.
14. Queueing theory studies waiting caused by random arrivals and service
Queues form when demand and service interact over time. Even when average service capacity exceeds average arrival rate, randomness can create waiting.
Queueing models study expected waiting time, system occupancy, utilisation and probabilities of delay under declared arrival and service assumptions.
The average rate alone is insufficient. Variability matters.
15. High utilisation can create nonlinear waiting
As service utilisation approaches full capacity, queues can grow sharply in many stochastic models because there is little spare capacity to absorb random bursts.
This explains why “keep every resource 100% busy” can be a poor system objective when responsiveness matters.
Idle capacity can function as a buffer against uncertainty rather than wasted capacity.
16. Little’s Law connects flow and inventory
Under broad steady-state conditions, Little’s Law relates average number in a stable system L, average throughput rate λ and average time in system W through L=λW.
The relation is powerful because it does not require the arrival and service distributions to take one narrow form, though the averages must be well defined under the operating conditions.
It links queue length, cycle time and throughput in one conservation relationship.
17. Inventory models balance shortage against holding
Inventory protects against uncertainty and lead time but costs money, space and obsolescence risk.
Operations research asks when to reorder, how much to order and how much safety stock is justified by the demand distribution and service target.
More stock is not automatically safer if the product becomes obsolete or ties up critical capital.
18. The economic order quantity shows a classical trade-off
In a simplified deterministic model with constant demand, fixed ordering cost and linear holding cost, the economic order quantity balances ordering frequency against inventory held.
The famous square-root formula is useful because it exposes the mechanism, but its assumptions should travel with it. Variable demand, quantity discounts and constrained storage require a different model.
19. Simulation is used when analytic structure becomes too complicated
Discrete-event simulation represents a system as events such as arrivals, departures, failures and repairs.
Simulation can compare policies in complex stochastic systems where closed-form analysis is unavailable.
But simulation estimates behaviour; it does not automatically optimise. A simulation must be combined with experimental design, search or optimisation if the goal is to find a best policy.
20. Random simulation needs uncertainty intervals
One simulation run is one random trajectory. Reliable conclusions require replication, variance estimates or other statistical controls.
A policy appearing better in one run may perform worse on average. Common random numbers and other variance-reduction techniques can make comparisons more efficient.
21. Decision trees organise sequential uncertainty
A decision tree alternates decision nodes and chance nodes. It clarifies which choices occur before uncertainty is resolved and which choices can adapt after information arrives.
This timing distinction matters. A plan that uses information unavailable at the decision time is not implementable.
22. Markov decision processes make policy dynamic
A Markov decision process combines states, actions, transition probabilities and rewards or costs. The decision changes the distribution of future states.
A policy maps states to actions. Dynamic programming can compare present reward with expected future consequence through Bellman-type recursions.
This creates a bridge between Stochastic Processes, optimisation and control.
23. Robust optimisation plans against a declared uncertainty set
If travel times, demand or capacities are uncertain, a nominal optimum can become infeasible.
Robust optimisation requires constraints to hold across an explicitly defined set of possible parameter values. This often sacrifices nominal performance to gain protection.
The uncertainty set is therefore part of the claim. A plan robust to normal demand variation is not automatically robust to a facility closure.
24. Stochastic programming allows scenario-dependent decisions
Stochastic programming represents uncertain scenarios with probabilities and allows some decisions to adapt after uncertainty is observed.
First-stage decisions occur before revelation; recourse decisions occur afterwards. This preserves the chronology of information.
The quality of the solution depends on the scenario model, not only the optimiser.
25. Multi-objective problems should not hide value trade-offs
Cost, speed, fairness, emissions and resilience may all matter. Combining them into one weighted score is possible, but the weights create value judgements.
Pareto-front analysis can instead show non-dominated alternatives and make trade-offs visible.
The mathematical result then supports a decision rather than pretending to replace the decision-maker’s values.
26. Sensitivity analysis asks what would change the recommendation
A solution depending on one exact demand forecast or cost coefficient may be fragile.
Sensitivity analysis varies parameters, studies shadow prices, inspects alternative optima and identifies thresholds where the chosen plan changes.
A robust operational recommendation should report these thresholds when they matter more than the last decimal place of the objective.
27. Decomposition makes enormous models manageable
Large systems often contain weakly coupled subproblems. Decomposition methods exploit that structure by solving smaller pieces and coordinating them through prices, cuts or master decisions.
The mathematical idea is organisational as well as computational: separate what can be solved locally, preserve the constraints that couple the pieces, and iterate until they agree sufficiently.
28. Heuristics can be useful without being proofs
Greedy methods, local search, tabu search, simulated annealing and other heuristics can find strong solutions to large hard problems.
Their usefulness does not require pretending they always find the global optimum. A good report distinguishes achieved feasible performance from proven optimality.
For operational decisions, a high-quality feasible answer available on time can be more useful than a proof arriving after the decision window closes.
29. A worked mechanism: exam timetabling
Suppose exams share students and rooms have limited capacity.
- Represent each exam as an entity.
- Create conflict edges between exams that cannot share a time slot.
- Create binary variables assigning each exam to a slot and room.
- Require exactly one assignment per exam.
- Forbid conflicting exams from sharing a slot.
- Respect room capacities.
- Optimise a declared criterion such as number of slots, student spacing or room use.
Each additional operational preference becomes a mathematical choice: hard constraint, soft penalty or post-solution criterion. That classification determines the result.
30. Common operations-research failure modes
- Optimising the wrong metric: improving the modelled objective while harming an omitted priority.
- Impossible fractional plans: relaxing indivisible decisions without restoring integrality.
- Average-only planning: ignoring variability in queues, demand and service.
- Forecast certainty: treating estimated parameters as exact.
- Solver worship: assuming a returned solution is globally optimal without status or gap evidence.
- Simulation certainty: treating one random run as expected behaviour.
- Timing errors: allowing a policy to use information that is not yet available.
- Implementation blindness: producing a mathematically feasible plan that cannot be executed by the real organisation.
31. Operations research as a mathematical machine
Decision Question → System Boundary → Variables/Constraints/Uncertainty → Optimisation or Simulation Model → Feasible Solution/Bounds → Sensitivity and Risk Check → Implementation → Measurement and Replanning.
The return step matters. A model should not end at “optimal.” It should state what needs to be measured after implementation and which deviations would require the plan to be recomputed.
32. What mastery looks like
- translate an operating problem into decision variables and constraints;
- recognise linear, integer, network, queueing and simulation structures;
- distinguish nominal, robust and stochastic formulations;
- use dual values and sensitivity without overextending them;
- separate feasible solutions from proven optima;
- model waiting and variability rather than averages alone;
- preserve information timing in dynamic decisions;
- treat heuristics honestly;
- return the mathematical plan to real operational constraints and monitoring.
33. Conclusion
Operations research works by turning complex operational decisions into inspectable models. Optimisation chooses among feasible plans. Graph theory structures networks. Probability and stochastic processes represent uncertainty. Queueing theory explains waiting. Simulation explores systems too complicated for closed forms. Sensitivity analysis reveals fragility. Implementation returns the mathematics to the organisation.
Mathematical optimisation asks for the best feasible choice. Operations research builds the larger decision system around that optimisation so the answer can survive contact with resources, uncertainty, time and operations.
How Mathematics Works | Batch 08
- How Mathematics Works | Control Theory
- Operations Research — this article
- How Mathematics Works | Cryptography
- How Mathematics Works | Mathematical Physics
Return to the Mathematics Learning Hub or the How Mathematics Works root.