What are we trying to improve?
Cost, service, throughput, risk, delay?
Multiple objectives may need explicit weights, priorities or trade-off analysis.
Side 20
Operations
Research
A study of turning messy operational choices into explicit models. Operations Research defines an objective, represents constraints, compares feasible alternatives and stress-tests the recommendation against uncertainty and real implementation limits.
objective→constraints→model→optimize→stress-test
06model families
05model steps
06failure checks
20Side
Optimization starts before the solver.
The largest errors often come from optimizing the wrong objective, omitting a constraint or treating an uncertain assumption as fixed.
What can actually be chosen?
Define decision variables.
Staffing, routing, quantities, schedules and assignments become variables the model can change.
What must remain true?
Capacity, budget, physics, policy, timing?
Constraints define the feasible region rather than merely penalizing undesirable outcomes.
Which parameters are inputs?
Measured, estimated or assumed?
Demand, processing time, cost and failure rates may carry substantial uncertainty.
Does the model represent the operational problem?
Compare with reality.
A mathematically optimal answer can still be operationally useless if the abstraction is wrong.
Canonical formmaximize or minimize f(x), subject to constraints on x
Find the best feasible choice under the model.
Different optimization families are defined by the structure of the objective, constraints and decision variables.
Linear objective, linear constraints.
Useful for allocation problems where contributions and resource use can be represented proportionally.
Some variables must be whole numbers.
Assignments, facility openings, yes/no decisions and indivisible units require discrete variables.
Relationships curve.
Costs, physics or returns can depend nonlinearly on decisions, changing both solution methods and local-optimum risks.
Break sequential problems into states.
Optimal decisions can be built recursively when the future depends on a compact representation of current state.
Inputs are uncertain.
Decisions are chosen while accounting explicitly for distributions or scenarios of future states.
There may be no single best answer.
Pareto-efficient solutions reveal trade-offs when improving one objective necessarily worsens another.
Optimal is conditional.
Every optimum is optimal only relative to the objective, constraints, data and assumptions encoded in the model.
Waiting is a system property.
Queueing models connect arrival patterns, service times, capacity and variability to congestion and delay.
Arrival
How does work enter?
Average arrival rate matters, but burstiness and time-of-day structure can matter just as much.
Service
How quickly is work processed?
Mean service time and its variability both affect waiting.
Capacity
How many servers operate?
Adding capacity can sharply reduce delay when utilization is high.
Discipline
Who gets served next?
First-come-first-served, priority rules and appointment systems produce different outcomes.
Utilization
How close is demand to capacity?
As utilization approaches full capacity, small variability can create disproportionately large queues.
Little’s LawL = λW
Average number in system = throughput rate × average time in system, under stable long-run conditions.
Average number in system = throughput rate × average time in system, under stable long-run conditions.
“Keep everyone busy” can be a bad queueing objective.
Very high utilization can maximize apparent resource use while producing severe waiting and fragile service levels.
Flows and stock decisions have structure.
Network and inventory models represent systems where location, connection, capacity and timing determine cost and service.
Network problems
Shortest path
Find the least-cost route.
Edges carry distance, time, risk or other costs.
Max flow
Move as much as possible.
Edge capacities constrain total flow from source to destination.
Assignment
Match resources to tasks.
Workers, machines or vehicles are allocated to jobs under compatibility and cost constraints.
Inventory problems
Order
How much?
Larger orders reduce ordering frequency but increase holding cost and exposure.
Reorder
When?
Lead time and demand uncertainty determine when replenishment should be triggered.
Safety
How much buffer?
Safety stock trades carrying cost against the risk and consequence of stockout.
When algebra becomes brittle, simulate the system.
Simulation generates possible system histories under specified rules, making it useful when interactions, randomness or nonlinearities make closed-form analysis difficult.
Define state
Specify the variables needed to describe the system at a moment in time.
Specify transitions
Encode how events, decisions and random draws change state.
Generate scenarios
Run many realizations rather than trusting one illustrative path.
Measure outcomes
Collect delay, cost, utilization, failures, service level or other relevant outputs.
Compare policies
Evaluate alternative decision rules under the same scenario conditions.
Validate
Check whether simulated behavior resembles the real system on known cases before trusting extrapolation.
Sample uncertain inputs repeatedly.
Useful for distributions of outcomes when formulas are difficult or risks are nonlinear.
Jump from event to event.
Useful for queues, factories, logistics and service systems where state changes at discrete moments.
Model interacting decision-makers.
Useful when aggregate outcomes emerge from heterogeneous local rules and interactions.
The best model answer may be the wrong operational answer.
Before implementation, test whether the recommendation survives uncertain inputs, model error and practical constraints that were simplified away.
Which assumptions move the answer?
Vary parameters to find where the recommended decision changes materially.
What if the environment changes?
Stress the model with demand spikes, supply failures, delays or cost changes.
Which constraint is expensive?
Marginal values reveal how much the objective could improve if a binding resource constraint were relaxed.
Prefer solutions that survive error.
A slightly worse nominal solution can be preferable if it performs acceptably across a much wider range of conditions.
Can humans and systems actually execute it?
Complex schedules may fail if they require unrealistic timing, perfect compliance or unavailable information.
Will behavior change after optimization?
People may adapt to incentives, queues and metrics, altering the system the model originally described.
Introduction to Operations ResearchHillier & Lieberman · broad OR foundation
Operations Research: Applications and AlgorithmsWayne Winston · modeling and methods
Introduction to Probability ModelsSheldon Ross · stochastic systems
Simulation Modeling and AnalysisAverill Law · simulation practice