THE GOLD STANDARD OF JUDGEMENT
Choose a rule while respecting how others will respond
A decision can change the incentives, routes and options of everyone who comes after it. This guide explains bilevel optimisation as a leader–follower model, then shows where tie-breaking, reformulation, uncertainty, fairness and solver limits change the judgment.
Choose a reading route
- Frame the hierarchy — identify commitment, sequence and control.
- Model response — build the follower before optimising the leader.
- Handle multiple responses — test tie-breaking and incentive substitution.
- Solve and verify — respect reformulation conditions and solver status.
- Stress and govern — examine uncertainty, fairness, pilots and transfer.
Use the chapter contents, jump to the worked leader–follower case, or return to The Gold Standard of Judgement. The wider routes are the Gold Standard of Education and the How X Works hub.
Chapter contents
Hierarchy and response · Chapters 1–4
Tie-breaking and incentives · Chapters 5–6
Solving and stress tests · Chapters 7–10
Governance and transfer · Chapters 11–12
The examples are fictional teaching models. They do not recommend a live price, policy or regulatory decision.
CHAPTER 1 OF 12 · Frame the hierarchy
1. A decision changes the problem others solve
Many policies, prices and designs are not followed by passive outcomes. A leader chooses a rule; people, firms or systems respond; the leader then receives the consequences of that response. Bilevel optimisation represents this hierarchy as one optimisation problem constrained by another.
The upper level may choose capacity, price, location or regulation. The lower level chooses a best response such as route, purchase, schedule or production. The leader cannot select both levels directly. Treating the follower’s decision as a controllable variable produces an attractive but infeasible plan.
Use the model only when the sequence and authority are real enough to matter. Some decisions are simultaneous, negotiated or repeated. Calling every interaction leader–follower can conceal feedback and power.
CHAPTER 2 OF 12 · Frame the hierarchy
2. Define the leader, follower and clocks
Name who decides first, what they observe and when each commitment becomes binding. A transport planner may expand a road before travellers choose routes. A retailer posts a price before customers choose quantities. A regulator sets a standard before firms redesign products.
The sequence can be informational rather than chronological. The follower may know the leader’s rule but not every implementation detail. The leader may anticipate an aggregate response rather than observe each individual.
Write the timing as a process diagram before writing equations. If the follower can respond before the leader’s action is fixed, or if the leader can revise immediately, the one-shot bilevel model may be the wrong representation.
CHAPTER 3 OF 12 · Model response
3. Build the follower problem first
The lower-level problem defines what the follower can do and what they prefer. It needs variables, constraints and an objective. A traveller chooses a feasible route and seeks lower travel time; a firm chooses production subject to technology and seeks profit.
Do not substitute the leader’s values into the follower. A public planner may value total emissions while a traveller values personal time. The difference is the point of the hierarchy.
Validate the follower model against observed or experimental response. If it predicts behaviour that people cannot perform or do not value, the upper-level optimum is built on fiction.
CHAPTER 4 OF 12 · Model response
4. Represent the leader objective and constraints
The leader evaluates the response together with its own decision. A planner might minimise investment cost plus congestion after travellers reroute. A platform might balance revenue, participation and fairness after users react.
Leader constraints can include budgets, legal limits, minimum service and safety. Put genuine boundaries in constraints rather than hoping a weighted objective will respect them.
State what the leader does not control: the follower’s tie-breaking, unobserved preferences, shocks and implementation errors. A complete mathematical expression can still omit decisive institutional limits.
CHAPTER 5 OF 12 · Handle multiple responses
5. Optimistic and pessimistic follower selection
A follower problem can have several equally good solutions. The optimistic formulation assumes the follower selects the one best for the leader. The pessimistic formulation assumes the one worst for the leader among follower optima.
Neither convention is automatically correct. The follower may use a stable but unknown tie-break, distribute across several options or behave heterogeneously. The gap between optimistic and pessimistic results is a useful measure of dependence on tie-breaking.
If the recommended leader action works only under favourable follower selection, report that fragility. Consider designing incentives or rules that remove the ambiguity.
CHAPTER 6 OF 12 · Handle multiple responses
6. Incentives can move, not command, behaviour
A tax, subsidy, price or priority changes the follower objective or feasible set. It does not guarantee a particular response unless the model and enforcement make that response uniquely optimal.
Strong incentives can create substitution. A fee on one route may shift traffic to another; a performance target may move effort from unmeasured tasks; a subsidy can attract participants who differ from the original population.
Trace the whole response space and affected groups. The leader’s improvement can be a transfer of cost rather than a net gain.
CHAPTER 7 OF 12 · Solve and verify
7. Reformulations require conditions
Some bilevel problems can be reformulated using optimality conditions for the follower, such as Karush–Kuhn–Tucker conditions in suitable continuous convex cases. Complementarity constraints or duality can convert the hierarchy into a single-level mathematical programme.
The conditions matter. Discrete follower choices, non-convexity or multiple optima can invalidate a convenient reformulation. A solver returning a solution does not prove the transformation represented the original problem.
Test small instances by enumeration. Compare the reformulated answer with direct follower optimisation for each candidate leader decision.
CHAPTER 8 OF 12 · Solve and verify
8. Complexity and solver status matter
Bilevel problems are often difficult. Mixed-integer choices and non-convex response can make global optimisation challenging. Local solutions, bounds and time limits must be reported.
Preserve solver version, tolerances, optimality gap, time limit, random seed where relevant and whether the follower response was solved exactly. A status such as feasible is not the same as proven optimal.
Use a simple baseline. If a transparent policy performs almost as well as a fragile complex optimum, the baseline may be the better judgment.
CHAPTER 9 OF 12 · Stress the model
9. Model uncertainty in response
Follower preferences, costs and constraints are estimated. Re-solve across plausible ranges. The leader decision may be stable even when predicted response changes, or it may sit on a cliff.
Scenario analysis should include behavioural adaptation, not only numerical noise. Followers may learn the rule, coordinate, exit the system or exploit loopholes.
Distinguish uncertainty about the response model from uncertainty about external conditions. They require different evidence and monitoring.
CHAPTER 10 OF 12 · Stress the model
10. Fairness, legitimacy and strategic power
A leader can design rules that extract value from followers. Mathematical optimality does not establish legitimacy. Ask who has voice, who bears transition cost and whether the response is voluntary, constrained or coerced.
Fairness constraints can protect minimum access or limit disparate burdens, but a narrow metric may miss procedural injustice. People may comply because alternatives are unavailable.
Use the model to expose power, not hide it. Publish the objective and constraints in language affected groups can challenge.
CHAPTER 11 OF 12 · Decide and govern
11. Pilot, monitor and update
A pilot tests whether the predicted response appears and whether unmodelled substitution emerges. Predefine measures, comparison periods and stop conditions. Avoid expanding because the dashboard moved in the desired direction once.
Monitor both leader outcomes and follower welfare. A pricing policy that reduces congestion by excluding lower-income users may meet one objective and fail the public purpose.
Update the model when preferences, technology or institutions change. A once-correct response function can become an engine of drift.
CHAPTER 12 OF 12 · Decide and govern
12. Transfer without copying the power structure
The bilevel structure transfers when one actor commits and another optimises in response. The values, constraints and legitimacy do not transfer automatically.
Move from transport to education and the follower may be a school, family or learner. The model must not assume people optimise a single score or that the authority may manipulate every incentive.
The gold standard is a decision whose hierarchy is explicit, response is validated, tie-breaking is tested, solver limits are visible and human consequences remain reviewable.
Worked example: opening study rooms while learners choose where to go
A fictional library can open capacity at North and East branches before an exam period. After seeing the capacities, learners choose the branch with the lower perceived crowding plus travel cost. The leader wants to reduce total crowding while spending no more than six capacity units.
| Leader choice | Cost | Predicted learner response | Total crowding score |
|---|---|---|---|
| Add 3 units at North | 3 | Many East learners switch to North | 22 |
| Add 3 units at East | 3 | North remains popular; East absorbs local demand | 20 |
| Add 3 at each branch | 6 | Demand balances if both remain equally attractive | 14 |
| Add 6 at North only | 6 | Strong switch to North creates a new queue | 18 |
Why the largest expansion can underperform
Adding all capacity at North looks attractive if demand is treated as fixed. Once learners reroute, the improved branch attracts a larger share and can recreate crowding. The follower response changes the upper-level objective.
Tie-breaking changes the result
Suppose the equal-capacity plan leaves learners indifferent between two branches. Under an optimistic assumption, they distribute evenly and the crowding score is 14. Under a pessimistic assumption, many choose the familiar North branch and the score rises to 17. The policy is still feasible, but the claimed gain depends on a behavioural tie-break.
Add a robust design choice
The library can publish live occupancy, reserve quiet seats at both branches and stagger workshop times. These design features reduce the sensitivity to one assumed response. They may be more valuable than solving the original model to another decimal place.
| Test | Observation | Judgment |
|---|---|---|
| Travel cost 20% higher | Fewer learners switch branches | Balanced expansion remains useful but gain narrows |
| North preference stronger | Optimistic tie-break fails | Add occupancy information or minimum East programme |
| Demand 15% above forecast | Both branches crowd | Capacity decision needs overflow rule |
| One branch closes temporarily | Original solution infeasible | Maintain a contingency rather than a static optimum |
Changed-context transfer
Move the model to a regulator setting a fee while firms choose compliance technology. The mathematics of sequence transfers, but the legitimacy test becomes stronger. Firms may pass costs to consumers, exit the market or contest the rule. The lower-level problem must represent lawful alternatives and affected groups, not only a profit function.
Practical verification
- Enumerate a small instance and compare every follower response with the reformulated solution.
- Solve optimistic and pessimistic versions when multiple follower optima are possible.
- Perturb response parameters and identify leader decisions that reverse.
- Compare the bilevel policy with a transparent fixed rule.
- Record which response was observed during a pilot and which model assumption it tests.
The model earns its place when it reveals incentive response and fragility. It loses value when the leader’s power or the follower’s humanity disappears inside the formulation.
Frequently asked questions
What makes a problem bilevel?
One decision-maker chooses while anticipating another optimisation problem whose solution constrains the first. Sequence and separate objectives are essential.
Is this the same as game theory?
It is closely related to Stackelberg leader–follower games. Bilevel optimisation often emphasises mathematical programmes and constraints; game models may include broader strategic structures.
Why can there be optimistic and pessimistic solutions?
The follower may have several equally optimal choices. The two conventions show how the leader outcome changes with favourable or unfavourable tie-breaking.
Can any bilevel model be reduced to one level?
No. Reformulations rely on conditions. Discrete, non-convex or multiple-response problems may require specialised methods.
Does the optimum predict behaviour?
Only if the follower model is valid. An optimisation solution is conditional on its behavioural assumptions.
When should a simpler rule win?
When performance is close and the simpler rule is more robust, explainable, lawful or easier to monitor.
Related reading and sources
Continue through eduKate
- Read The Gold Standard of Judgement for the root framework.
- Use Goal Programming, Priority Levels and Deviation Variables for aspiration-based decisions.
- Use Multi-Choice Knapsack, Shadow Prices and Portfolio Constraints for a single-level portfolio problem.
- Use Robust Decisions and the Assumption That Reverses a Choice for sensitivity.
Technical references
- Indian Institute of Management Ahmedabad: Bilevel Optimization—Applications, Models and Solution Approaches.
- Carnegie Mellon University discopt notebook: Bilevel Optimization and Stackelberg games.
- University of Southampton: Multilevel Optimisation.
- IIMA tutorial: Evolutionary Bilevel Optimization.
Return to the Gold Standard of Education or the How X Works hub.
Advanced practice laboratory: a fee that changes the follower’s choice
Consider a fictional campus with two shuttle routes. The leader sets a peak fee on Route A to reduce crowding. Each traveller then chooses A, Route B or a longer walking route. The leader values lower total delay and affordable access. Travellers value their own time and fee. This is a leader–follower structure because the fee changes the choice problem faced by travellers.
Enumerate before optimising
Use a small demand model first. At fee 0, 80 travellers choose A, 20 choose B and none walk. At fee 1, 60 choose A, 35 choose B and 5 walk. At fee 2, 42 choose A, 43 choose B and 15 walk. At fee 3, 30 choose A, 40 choose B and 30 walk. These are fictional response estimates, not a real transport model.
| Fee | A riders | B riders | Walkers | Leader delay score | Access burden |
|---|---|---|---|---|---|
| 0 | 80 | 20 | 0 | 38 | Low monetary burden; high A crowding |
| 1 | 60 | 35 | 5 | 25 | Moderate; small walking shift |
| 2 | 42 | 43 | 15 | 18 | Higher burden; routes balanced |
| 3 | 30 | 40 | 30 | 17 | Little delay gain; large access burden |
If the leader minimises delay alone, fee 3 appears best. If the objective includes walking burden and affordability, fee 2 may dominate. The follower response exposes a trade-off; it does not choose the public value.
Make the lower-level choice explicit
A traveller chooses the option with the lowest perceived generalised cost: travel time, waiting, fee, reliability and accessibility. Different travellers have different values and constraints. One representative follower can hide the fact that a fee changes choices mainly for people with lower ability to pay.
Segment only when evidence and purpose justify it. At minimum, test travellers who cannot walk the longer route, who connect to fixed schedules or who receive concession support. Feasible options are as important as preferences.
Multiple best responses
At fee 2, some travellers may be indifferent between A and B. An optimistic leader model spreads them to minimise crowding. A pessimistic model sends them all to the already busier route. A behavioural model may use historical habit. Solve all three. If the leader’s decision changes, the tie-break is a policy risk.
Why KKT reformulation may fail
If follower choice is continuous, convex and regular, optimality conditions can help replace the lower-level problem. Here the choices are discrete and many travellers act separately. A smooth approximation may be useful, but it is not the same problem. Validate it against enumeration or simulation. Complementarity conditions written correctly can still model the wrong behaviour.
Add constraints the objective must not buy away
- Maintain a no-fee accessible option for travellers who cannot use the walking route.
- Cap the maximum expected wait on each route.
- Limit the share of lower-income users facing the peak fee without a concession.
- Keep emergency access and staff transport outside the incentive experiment.
- Require a review if displacement raises safety concerns on the walking route.
Test strategic adaptation
Travellers may shift departure times, share passes, crowd the period just outside the peak or stop using the service. Route B may alter its schedule after observing demand. A one-shot model misses these adaptations. Add scenarios or move to a repeated game when responses feed back across periods.
Compare a transparent rule
A simple alternative opens an extra vehicle when Route A occupancy exceeds a threshold. It may cost more but avoid distributional burden. Compare total delay, operating cost, access, explainability and failure modes. The mathematically optimal fee should not be compared only with doing nothing.
Pilot design
Predefine a short pilot, notice, exemptions, data minimisation, stop conditions and a baseline period. Measure route load, waiting, walking displacement, missed connections, complaints and distributional effects. Do not tune the fee daily until the desired result appears; that destroys the evaluation.
Decision after the pilot
Suppose fee 2 balances routes but walking rises sharply among users without flexible schedules. The leader can reduce the fee, improve Route B, add a concession or reject pricing. The model has succeeded if it made response and burden visible—even if the chosen policy is not its original optimum.
Changed-context transfer: capacity procurement
A buyer announces contract terms and suppliers decide whether and how much to bid. The leader–follower structure transfers, but market entry, collusion, legal procurement duties and long-term supplier capacity must be represented. A price-response curve from shuttle users cannot be reused. Transfer the sequence, then rebuild objectives, constraints and evidence.
Verification record
- Leader decision, objective and non-compensable constraints.
- Follower options, information, objective and validated response data.
- Optimistic, pessimistic and behavioural tie-breaking results.
- Reformulation assumptions and small-instance checks.
- Solver status, gap, tolerances and alternative near-optimal policies.
- Pilot outcomes, distributional effects and reasons for the final decision.
Model card: what must accompany a leader–follower recommendation
A bilevel result is difficult to interpret without its institutional story. Attach a model card that states who commits, who responds, what information each side has, whether the interaction repeats and which choices are outside the model. This prevents a later reader from treating a mathematical hierarchy as a legal or moral entitlement to lead.
Response evidence
Describe how the follower objective was estimated: revealed choices, survey, experiment, engineering model or assumed rationality. State sample, period, exclusions and known changes since collection. If the follower is an organisation, identify internal decision rules rather than assigning it a single preference without evidence.
Solution evidence
- Original upper- and lower-level formulations.
- Any KKT, duality, big-M or complementarity reformulation and the conditions it needs.
- Bounds chosen for big-M constants and tests showing they do not cut off valid solutions.
- Optimistic or pessimistic convention for multiple follower optima.
- Solver status, best bound, incumbent objective and time limit.
- Independent evaluation of the follower response at the reported leader decision.
Large big-M values can weaken computation; small values can remove feasible solutions. A successful solver termination cannot validate an unjustified constant. Derive bounds from the domain where possible and test them.
Near-optimal policies
List several leader decisions within a meaningful tolerance of the best objective. Compare interpretability, transition cost, fairness, resilience and monitoring burden. A one-per-cent objective gain may not justify a policy that is hard to explain or sensitive to tie-breaking.
Counterfactual limits
The model predicts response under its specified options and preferences. It does not automatically estimate what people would do if they rejected the institution, coordinated politically or changed the rules. State which counterfactuals are supported and which are outside scope.
Governance after deployment
Assign owners for response drift, constraint breaches and distributional complaints. Publish a review date and exit condition. If observed behaviour persistently falls outside the prediction interval or new follower options appear, suspend automatic use and revalidate.
Final decision question
Before approval, ask: if the followers read the model and understood how their behaviour was represented, could the leader explain the objective, constraints and safeguards without hiding behind the solver? That test does not settle legitimacy, but it exposes when optimisation has replaced public reasoning.
Failure laboratory: five ways a correct solve gives a poor decision
The follower is modelled as more rational than the institution
A firm may optimise profit, but internal budgets, contracts and approval cycles delay response. A mathematically best action can be organisationally unavailable. Add implementation constraints or treat the model as a long-run scenario.
The leader cannot commit
Stackelberg logic assumes the leader’s move is credible. If the leader can reverse the fee after seeing response, followers may wait, lobby or choose strategically. Model repeated interaction or strengthen the commitment mechanism.
The response data come from another regime
Preferences estimated before a policy change may not describe behaviour after it. A fee can change norms, information and available alternatives. Use a pilot and update rather than treating historical elasticity as timeless.
The solver selects a fragile corner
An optimum may rely on a follower being exactly indifferent. Small numerical or behavioural changes then produce a different response. Report distance to the switching boundary and prefer an interior, slightly suboptimal policy when it is materially more stable.
The objective omits the harmed party
A low congestion score can be purchased through exclusion. Add distributional measures and hard access constraints before solving. Do not append an ethical paragraph after an objective that already made the relevant people invisible.
Recovery exercise
For each failure, define a detection signal and repair: compare predicted with observed response; test commitment; re-estimate after regime change; run local sensitivity; and add affected-party review. The best model governance asks not only whether the code works, but which assumption will fail first and how the institution will notice.
Decision handoff
The final brief should show one recommended policy, at least two near-optimal alternatives, the follower responses they assume, the tie-break convention, the groups that bear cost, and the trigger for reconsideration. This turns an optimisation output into a reviewable judgment.