Optimisation is the search for the best feasible solution to a stated objective while respecting stated constraints.
In one line: optimisation works by defining what “better” means, choosing the decisions that can change, representing the limits that must be respected, comparing feasible alternatives, selecting the best solution under the model, then testing whether that solution remains useful when uncertainty, human values and real-world conditions return.
Evidence boundary: Mathematical optimisation has precise technical definitions. The NEOS Guide describes an optimisation problem as finding the best solution among feasible solutions and models it through an objective, variables and constraints. This article explains that core while also showing where real-world decisions exceed the mathematical model. “Optimal” always means optimal for the objective and constraints represented; it does not automatically mean morally best, socially fair or globally wise.
Optimisation sounds like “make everything better.”
That is not precise enough. A system cannot optimise until it knows what it is trying to maximise or minimise and which limits it cannot ignore.
The objective creates direction. The constraints create reality.
PHASE 4++++ · DEFINITIVE OPTIMISATION MAP
How to Read This Optimisation Guide
Optimisation is disciplined choice inside a represented feasible world. It can search enormous spaces and identify solutions that outperform human intuition, but it cannot decide by mathematics alone which objective deserves to be optimised, which human values must become hard constraints, whose burden may be traded away, or whether the model has represented enough of reality for its optimum to deserve implementation.
The invariant optimisation chain
Real decision problem and receiver → objective or objectives → decision variables → constraints and protected floors → feasible set → data, model and uncertainty → solver or search method → candidate solutions → optimum or trade-off frontier under the model → sensitivity, robustness and distributional test → authorised choice → implementation → observed outcomes → objective, constraint or model update.
The final world return matters because an optimiser solves the represented problem. Implementation reveals omitted behaviour, new bottlenecks, changing demand, gaming, hidden costs and values that were not adequately encoded.
What this article owns
This article owns the mechanism for searching a feasible solution space for the best represented solution under one or more objectives and constraints.
Decision-Making owns goals, values, authority and commitment. Capacity and Scarcity own important resource constraints. Measurement and Data own inputs. Forecasting and Uncertainty own future and model limits. Sensitivity Analysis tests fragile optima. Control Systems can use optimisation repeatedly inside a live regulatory loop. Safety and Quality can define protected constraints that performance optimisation may not trade away.
The receiver, value and robustness test
A system-wide optimum can improve the average while concentrating delay, danger, cost or lost access on a smaller group. A mathematically superior plan can also sit so close to a constraint that ordinary forecast error makes it infeasible in practice.
The practical question is: Best according to whose objective, under which protected constraints, for which receivers, and does the solution remain acceptable when the uncertain world moves away from the estimated values?
Four distinctions that prevent optimisation from becoming authority
- Optimal under the model is not best in every sense. Omitted objectives, constraints and human consequences remain omitted from the mathematical solution.
- Weights are not neutral facts. Combining objectives into one score encodes judgements about how much one outcome can compensate for another.
- Local optimum is not global optimum. The guarantee depends on the problem structure and method; a good nearby solution may not be the best feasible solution overall.
- Mathematical recommendation is not legitimate authority. A solver can identify a candidate route; accountable people and institutions remain responsible for whether it may be implemented.
Clickable deep routes
- How Decision-Making Works — objectives, values, thresholds, reversibility and authority around the mathematical recommendation.
- How Capacity Works and How Scarcity Works — the resource and flow constraints that shape the feasible set.
- How Measurement Works and How Data Works — where costs, capacities, demand and other inputs acquire meaning and limits.
- How Uncertainty Works and How Sensitivity Analysis Works — whether the optimum survives plausible changes in assumptions and inputs.
- How Resilience Works — why margin and graceful degradation can outperform knife-edge nominal perfection.
- How Control Systems Work — when optimisation becomes one action-selection component inside a repeated live feedback loop.
- How X Works Master Hub — return to the wider mechanism library.
Definitive optimisation does not ask only whether the solver found the best number. It asks whether the right problem was represented, the protected floors remained inviolable, uncertainty was strong enough to change the choice, burden was visible, authority was legitimate, and the world returned the improvement the objective claimed to value.
What Is an Optimisation Problem?
A useful optimisation chain is:
Decision problem → objective → decision variables → constraints → feasible set → model/algorithm → candidate solutions → optimum under model → sensitivity/robustness → real-world verification → update objective, constraints or model.
1. The Objective Defines What “Better” Means
An objective is a quantity or criterion the system tries to improve.
Minimise travel time. Minimise cost. Maximise output. Maximise expected learning gain. Minimise energy use. Maximise reliability.
The NEOS Guide describes the objective as a quantitative measure of system performance to minimise or maximise.
The objective matters because whatever is omitted from it can become invisible to the optimiser.
2. Decision Variables Are the Things the Optimiser Can Change
Decision variables represent choices under the decision-maker’s control.
How many units should each factory produce? Which route should a vehicle take? How should staff be scheduled? How much study time should be allocated to each topic?
If a factor cannot be changed, it usually enters the model as data, an assumption or a constraint rather than a decision variable.
3. Constraints Define the Feasible World
Constraints describe what solutions are allowed.
Budget cannot exceed available money. Staffing cannot exceed available people. A vehicle cannot carry more than its safe load. A student cannot allocate 30 hours of revision inside a 24-hour day.
Constraints protect the model from recommending impossible solutions.
They can also represent law, safety, fairness or policy boundaries when those requirements are explicit enough to model.
4. The Feasible Set Contains the Solutions That Respect the Constraints
Once the constraints are applied, many imaginable solutions disappear.
The remaining possibilities form the feasible set.
The optimiser’s job is not to choose from every conceivable action. It is to find the best candidate inside this allowed region.
5. Infeasibility Is Useful Information
Sometimes no solution satisfies every constraint.
This is not merely a software failure. It can reveal that the real requirements are mutually incompatible.
If a project must be completed tomorrow, cost nothing, use no additional staff and meet a high safety threshold, the constraints may simply leave no feasible plan.
The correct response is to revisit the constraints or resources rather than fabricate a solution.
6. Different Problem Types Need Different Methods
Optimisation problems differ in important ways.
Variables may be continuous or discrete. Relationships may be linear or nonlinear. Constraints may be simple or complex. Uncertainty may be ignored, represented probabilistically or explored through scenarios.
The NEOS Guide emphasises classifying the problem because algorithms are designed for different mathematical structures.
7. An Optimal Solution Is Still Conditional on the Model
A solver can find the mathematically optimal answer to the wrong problem.
If a delivery model minimises distance but ignores driver hours, safety or access restrictions, its “optimal route” may be unusable. If a school optimises only examination averages, it may neglect students or capabilities that the metric does not reward.
Optimality is therefore conditional: best under these represented objectives, variables, constraints, data and assumptions.
8. Local and Global Optima Are Different
In some problems, several peaks or valleys exist.
A local optimum is better than nearby alternatives. A global optimum is best across the entire feasible set.
Some algorithms can guarantee global optimality for particular problem classes; others may return strong local solutions without proving no better solution exists elsewhere.
The certainty of the solution should match what the method can actually establish.
9. Multiple Objectives Create Trade-Offs
Real systems rarely care about one outcome only.
A transport plan may care about cost, travel time, emissions, reliability and accessibility. A hospital may care about waiting time, safety, staff workload and cost. A learner may care about marks, understanding, sleep and long-term retention.
These objectives can conflict.
Multi-objective optimisation can expose trade-off frontiers rather than pretending one solution improves everything simultaneously.
Singapore turns this multi-objective problem into an applied systems question because housing, transport, jobs, schools, healthcare, land, cost, access and resilience cannot all be optimised independently. Continue to What Makes Singapore Work | Systems Fit for the question of whether improvements in one system still fit the capacities, timing and receiver needs of the others; this page remains the general mechanism owner for optimisation.
10. Weighting Objectives Encodes Values
If several objectives are combined into one score, the weights determine how much one outcome can compensate for another.
Those weights are not purely mathematical facts. They encode preferences, policy choices or ethical judgements.
When consequences are important, weighting should be visible enough for affected people to understand what the system is trading away.
11. Constraints Can Protect Values Better Than Weights
Some outcomes should not simply be traded for enough improvement elsewhere.
Safety limits, legal rights, minimum service levels and hard budget ceilings may be better represented as constraints than as weak penalties in a weighted score.
This modelling choice matters because a penalty can sometimes be “worth paying” inside the objective, while a hard constraint forbids the violation.
12. Optimisation Needs Reliable Data
Travel times, costs, capacities, demand, failure rates and other inputs determine the recommendation.
If the data are stale, biased or poorly measured, the optimisation can become precisely wrong.
For the evidence-generation layer, see How Measurement Works and How Evidence Works.
13. Uncertainty Changes the Best Solution
A solution optimised for one exact forecast may perform badly when demand, travel time or prices differ.
Stochastic, scenario-based or robust optimisation methods can account for uncertainty in different ways.
The broader principle is accessible without advanced mathematics: do not optimise as if uncertain inputs were known perfectly when the decision is sensitive to them.
14. Robustness Can Be Better Than Fragile Perfection
The mathematically best solution at the estimated values may sit very close to a constraint.
A small forecasting error can then make the plan fail.
A slightly less efficient solution with more margin may perform better across a range of plausible conditions.
Optimisation therefore connects to How Resilience Works: maximum nominal performance and robust real-world performance are not always the same.
15. Sensitivity Analysis Shows What Really Drives the Answer
Change one important assumption or parameter and see how the solution changes.
If a tiny change in fuel price completely changes the route plan, the recommendation is sensitive. If many plausible input changes leave the same solution preferred, confidence can increase.
Sensitivity analysis helps distinguish robust recommendations from knife-edge optima.
16. Proxy Objectives Can Be Gamed
Often the true goal is hard to measure, so a proxy is optimised instead.
A call centre may optimise average handling time when the deeper goal is useful resolution. A school may optimise test scores when the deeper goal includes understanding, transfer and long-term capability.
Once people and systems adapt to the metric, the proxy can separate from the underlying purpose.
The cure is not “never measure.” It is to keep the objective connected to receiver outcomes and monitor for gaming and neglected dimensions.
17. Optimisation Can Redistribute Burden
A system-wide optimum can hide who pays the cost.
A transport route that minimises total travel time might worsen access for a small community. A staffing plan that minimises labour cost may create unacceptable fatigue for a subset of workers.
Distributional effects should therefore be measured or constrained when human consequences matter.
18. Real-World Verification Comes After the Solver
Optimisation produces a recommendation inside a representation of reality.
Implementation tests whether the omitted details matter.
Did queues fall? Did cost actually decrease? Did users change behaviour? Did a new bottleneck appear? Did the plan create a safety or fairness problem the model failed to represent?
Strong optimisation returns to evidence after action and updates the model.
19. Education Needs Optimisation Without Treating the Child as a Machine
Students face genuine optimisation problems: limited revision time, several subjects, different weaknesses and changing examination dates.
But the objective should not collapse the learner into one score. Sleep, health, understanding, motivation, future options and transfer can matter alongside immediate marks.
A strong study plan optimises the plan around the learner’s current evidence while leaving the learner more capable and autonomous, not merely more efficiently controlled.
The Whole Optimisation Chain
Real problem → objective → decision variables → constraints → feasible set → data and model → algorithm → candidate solutions → optimum under model → sensitivity + uncertainty + distributional check → implementation → observed outcome → update model/objective/constraints → optimise again if useful.
A Useful Metaphor: Optimisation Is Choosing the Best Route on a Map
“Best route” depends on the objective.
The shortest route may not be fastest. The fastest may have tolls. The cheapest may be unreliable. A route unsuitable for a wheelchair, heavy vehicle or flood condition is not feasible for that receiver.
The map can optimise only what it represents. Reality still gets the final vote.
Optimisation at Three Zoom Levels
Micro: one decision
What objective is being improved and which constraints make a solution feasible?
Meso: one operating system
Do data, trade-offs, uncertainty and bottlenecks make the chosen optimum robust enough for real operation?
Macro: institution or society
Who chose the objective, whose values became constraints, who receives the gains and who carries the burden?
How Optimisation Fails
- Wrong objective: the system optimises a proxy that does not protect the real purpose.
- Missing constraint: the recommended solution is mathematically feasible but practically, legally or ethically unacceptable.
- Bad data precision: uncertain inputs produce an exact-looking optimum.
- Fragile optimum: tiny input changes make the solution fail.
- Local-optimum confusion: a strong nearby solution is presented as globally best without proof.
- Distribution blindness: total performance improves while one group absorbs disproportionate cost.
- Model capture: people begin serving the optimisation score instead of the underlying receiver.
- No world return: the recommendation is never checked after implementation.
How Optimisation Is Strengthened
Define the real objective with the receiver visible. Separate decision variables from assumptions. Encode hard boundaries as constraints where appropriate. Check feasibility before chasing optimality. Use trustworthy data. Test uncertainty and sensitivity. Compare multiple objectives openly. Inspect distributional effects. Prefer robust solutions when fragile precision is dangerous. Implement, observe and let reality update the next model.
What Parents and Students Should Notice
- What exactly is the study plan trying to optimise?
- Which constraints are real: time, examination dates, sleep, prerequisite knowledge or attention?
- Is a score being used as a proxy for deeper capability?
- Would a small change in assumptions completely change the plan?
- Is the “efficient” plan leaving enough resilience for illness or unexpected work?
- Who chose the weights among marks, understanding, rest and long-term retention?
- After the plan runs, is actual performance used to update the next allocation?
Continue Through eduKateSG
- How Decision-Making Works
- How Capacity Works
- How Scarcity Works
- How Forecasting Works
- How Resilience Works
Evidence and Further Reading
The NEOS Guide to Optimization defines optimisation as finding the best solution among feasible solutions and introduces the three central modelling elements used in this article: the objective, variables and constraints.
The NEOS Guide’s problem-types overview explains why constrained/unconstrained, continuous/discrete, deterministic/stochastic and other mathematical structures require different optimisation approaches rather than one universal algorithm.
Frequently Asked Questions
Does optimal mean best in every sense?
No. It means best according to the represented objective while satisfying the represented constraints. Important human, ethical or practical factors omitted from the model remain omitted from its claim of optimality.
What happens if there is no feasible solution?
The constraints cannot all be satisfied simultaneously. The decision-maker must add resources, change the objective, relax an appropriate constraint or redesign the problem rather than pretending a feasible answer exists.
Why can a slightly worse mathematical solution be better in practice?
Because a solution with more margin may remain effective across uncertain conditions, while a knife-edge optimum can fail after a small forecasting or measurement error. Robustness can have real value that the original objective did not fully represent.
Final compression: optimisation is disciplined trade-off, not magic. It can find the best feasible solution inside a model, but wisdom still requires checking whether the objective was worth optimising, the constraints protected what mattered and the real world behaved enough like the model for the answer to deserve use.