In the previous lesson, we looked at single-objective optimisation.

The idea was straightforward:

Choose the feasible solution that performs best according to one objective.

For example:

[ \min \text{cost} ]

or:

[ \max \text{utility} ]

But most important real-world decisions are not that simple.

We usually care about several things at once.

A hospital might want to:

maximise patient benefit
minimise waiting times
minimise cost
protect vulnerable patients
maintain resilience

An electricity system might want to:

minimise cost
minimise emissions
maximise reliability
minimise network congestion
distribute resources fairly

These objectives can conflict.

That gives us multi-objective optimisation.

More than one definition of "better"

Suppose we have two objectives:

[ f_1(x) ]

and:

[ f_2(x) ]

We might want to:

[ \min f_1(x) ]

and simultaneously:

[ \min f_2(x) ]

For example:

[ f_1(x)=\text{cost} ]

and:

[ f_2(x)=\text{emissions} ]

The problem is that the cheapest solution may not be the cleanest.

And the cleanest solution may not be the cheapest.

There may therefore be no single solution that is best according to both objectives.

A simple example

Imagine three ways of generating electricity.

Option	Cost	Emissions
A	£40	100
B	£60	50
C	£90	10

If we optimise only:

[ \min \text{cost} ]

we choose:

[ A ]

If we optimise only:

[ \min \text{emissions} ]

we choose:

[ C ]

But perhaps we care about both.

Now the answer is less obvious.

Option B might represent a compromise.

But whether that compromise is desirable depends on how we value cost relative to emissions.

There may be no single optimum

In single-objective optimisation, we can usually rank solutions according to one number.

If:

[ J(A)<J(B) ]

then A is better.

With multiple objectives, things become more complicated.

Suppose:

[ A=(40,100) ]

and:

[ C=(90,10) ]

where the first number is cost and the second is emissions.

A is better on cost.

C is better on emissions.

Neither is universally better.

This means the solutions are incomparable without additional preferences.

Dominance

Multi-objective optimisation introduces an important idea:

dominance

Suppose we are minimising both cost and emissions.

Consider:

Option	Cost	Emissions
A	£40	100
B	£60	50
C	£90	10
D	£70	80

Compare B and D.

B has:

[ 60 < 70 ]

and:

[ 50 < 80 ]

So B is better than D on both objectives.

We say:

B dominates D.

There is no reason to choose D if B is feasible.

Pareto dominance

More formally, suppose we are minimising several objectives:

[ f_1(x),f_2(x),\ldots,f_m(x) ]

Solution (x_a) dominates (x_b) if:

[ f_i(x_a)\leq f_i(x_b) ]

for every objective (i),

and it is strictly better on at least one objective.

In simple terms:

A solution dominates another if it is no worse at everything and better at something.

This is called Pareto dominance.

Pareto optimality

Now return to:

Option	Cost	Emissions
A	£40	100
B	£60	50
C	£90	10

None of these options dominates another.

A is cheapest but dirtiest.

C is cleanest but most expensive.

B sits between them.

All three may therefore be Pareto optimal.

A Pareto-optimal solution is one where:

You cannot improve one objective without making at least one other objective worse.

The Pareto frontier

If we plot many possible solutions, we may find a boundary containing the best trade-offs.

This is called the:

Pareto frontier

or:

Pareto front

Imagine cost on one axis and emissions on another.

Many possible solutions might exist:

[ x_1,x_2,\ldots,x_n ]

Some are clearly bad.

They are:

expensive
polluting

The optimiser can reveal the choices

This is an important difference from single-objective optimisation.

A multi-objective optimiser does not necessarily tell us:

Here is the answer.

Instead, it may tell us:

Here are the best available trade-offs.

That is often much more useful.

The optimiser can reveal the structure of the decision.

Humans, institutions or other decision mechanisms can then decide which trade-off is acceptable.

Multi-objective optimisation separates possibility from preference

This gives us a useful distinction.

The optimisation problem can tell us:

What combinations are achievable?

Then preferences can tell us:

Which achievable combination do we want?

These are different questions.

For example, engineering may tell us that we can achieve:

Option	Cost	Reliability
A	£1bn	99.0%
B	£1.2bn	99.9%
C	£1.8bn	99.99%

The optimisation exposes the possibilities.

But engineering alone cannot necessarily tell society:

how much additional reliability is worth paying for.

That requires a value judgement.

One approach: weighted sums

A common way to handle multiple objectives is to combine them into one.

Suppose we care about:

[ C(x)=\text{cost} ]

and:

[ E(x)=\text{emissions} ]

We can create:

[ J(x)=\alpha C(x)+\beta E(x) ]

and solve:

[ \min J(x) ]

where:

[ \alpha,\beta ]

represent the relative importance of the objectives.

Technically, we have converted the problem back into a single-objective optimisation problem.

The weights change the answer

Suppose:

[ J=C+0.1E ]

Cost receives relatively more importance.

We may choose a cheap, relatively polluting solution.

Now suppose:

[ J=C+100E ]

Emissions become extremely important.

The optimiser may accept much higher financial cost to reduce them.

The algorithm has not changed.

The feasible region has not changed.

The weights changed.

And therefore the decision changed.

Weights encode values

This is why weights deserve attention.

A model might appear highly mathematical:

[ J= 0.6f_1 + 0.3f_2 + 0.1f_3 ]

But where did:

[ 0.6,\ 0.3,\ 0.1 ]

come from?

Perhaps they were:

chosen by policymakers
estimated from surveys
inferred from behaviour
selected by engineers
inherited from an old model
chosen simply because they produced reasonable-looking results

The optimiser cannot tell us whether those weights are morally or politically correct.

It simply uses them.

Normalisation matters

There is another problem.

Suppose:

[ f_1=\text{cost in pounds} ]

and:

[ f_2=\text{emissions in tonnes} ]

Cost might have values around:

[ 1,000,000 ]

while emissions might have values around:

[ 100 ]

If we simply write:

[ J=f_1+f_2 ]

then cost will numerically dominate the objective.

That does not necessarily mean cost is more important.

It may simply have larger units.

Normalising objectives

One solution is to put objectives onto comparable scales.

For example:

[ \hat{f}_1

\frac{f_1-f_1^{\min}} {f_1^{\max}-f_1^{\min}} ]

and similarly:

[ \hat{f}_2

\frac{f_2-f_2^{\min}} {f_2^{\max}-f_2^{\min}} ]

Now both might range approximately between:

[ 0 ]

and:

[ 1 ]

We can then write:

[ J= \alpha \hat{f}_1 + \beta \hat{f}_2 ]

This makes the weights easier to interpret.

But the choice of normalisation still matters.

Another approach: turn objectives into constraints

Suppose our two goals are:

minimise cost
minimise emissions

Instead of combining them, we might say:

[ \min C(x) ]

subject to:

[ E(x)\leq E^{\max} ]

Now emissions become a constraint.

We are asking:

What is the cheapest solution that meets our emissions target?

This can be much easier to explain.

The reverse formulation

We could equally write:

[ \min E(x) ]

subject to:

[ C(x)\leq C^{\max} ]

Now we ask:

What is the cleanest solution we can afford within this budget?

Same two concerns.

Different framing.

This shows again that:

How we formulate the problem influences the answer.

The epsilon-constraint method

This idea can be generalised.

Suppose we have:

[ \min f_1(x) ]

while requiring:

[ f_2(x)\leq \epsilon ]

We can vary:

[ \epsilon ]

and repeatedly solve the problem.

For example:

[ E\leq100 ]

then:

[ E\leq80 ]

then:

[ E\leq60 ]

and so on.

Each solution reveals how much the primary objective changes as we tighten the secondary requirement.

This is one way of exploring the Pareto frontier.

Hard requirements versus preferences

This raises a deeper design question.

Suppose we care about:

cost
safety

Should we write:

[ J= \alpha C+\beta S ]

and trade one against the other?

Perhaps not.

We might instead say:

[ \min C ]

subject to:

[ S\geq S^{\min} ]

Now safety is not something the optimiser can trade away for enough money.

It is a requirement.

This distinction matters enormously.

Some things should not necessarily be traded

Imagine an autonomous vehicle choosing between:

journey time
passenger comfort
safety

Journey time and comfort might reasonably be traded.

Safety may need a hard minimum.

Similarly, a public system might treat:

legal rights
physical safety
minimum access
essential services

as constraints rather than weighted preferences.

Multi-objective optimisation therefore forces us to ask:

What may be traded, and what must be protected?

Electricity systems are inherently multi-objective

Consider operating an electricity system.

We might care about:

generation cost
carbon emissions
reliability
voltage quality
congestion
renewable utilisation
consumer affordability
fairness
resilience

These are not naturally identical objectives.

For example, maximising renewable utilisation might sometimes increase:

network reinforcement requirements
storage requirements
curtailment elsewhere

Maximising reliability might require additional infrastructure.

Minimising short-term cost might reduce long-term resilience.

The real system contains trade-offs.

Healthcare is also multi-objective

Imagine a hospital allocating scarce operating-theatre capacity.

Possible objectives include:

maximise health benefit
minimise waiting time
prioritise urgent cases
reduce inequality
minimise cost
maximise utilisation

A system that optimises only:

[ \max \text{total health benefit} ]

might allocate resources differently from one that prioritises:

[ \min \text{worst waiting time} ]

Neither objective is automatically absurd.

They represent different values.

Transport systems contain competing objectives

A city might want to:

minimise travel time
minimise emissions
maximise accessibility
reduce congestion
protect pedestrian safety
improve public transport
minimise infrastructure cost

Improving one can affect another.

For example, a road expansion might initially reduce:

[ \text{vehicle travel time} ]

but encourage more driving and increase:

[ \text{emissions} ]

A transport system cannot be understood properly through one metric alone.

AI systems are often multi-objective too

Consider a large AI model.

We might care about:

accuracy
latency
computational cost
energy consumption
robustness
safety
interpretability

A larger model might improve accuracy but require:

[ 10\times ]

the compute.

A smaller model might be slightly less accurate but:

much faster
cheaper
easier to deploy

Which is better?

There is no answer until we define the context.

Recommendation systems provide another example

Suppose a platform recommends content.

It might optimise:

[ \max \text{engagement} ]

But the platform may also care about:

user satisfaction
diversity
safety
long-term retention
misinformation
creator fairness

Maximising engagement alone may produce very different behaviour from balancing several objectives.

This is one reason seemingly simple AI products become system-design problems.

Short-term and long-term objectives can conflict

Multi-objective problems do not only involve different types of value.

They can also involve different timescales.

A company might want to:

[ \max \text{profit today} ]

and:

[ \max \text{long-term viability} ]

These can conflict.

Cutting maintenance may increase today's profit.

But it may damage tomorrow's infrastructure.

Similarly:

[ \min \text{current electricity cost} ]

may conflict with:

[ \max \text{future system resilience} ]

This means time itself can create competing objectives.

Individual and system objectives can conflict

Suppose every household wants:

[ \min \text{my electricity bill} ]

Every generator wants:

[ \max \text{my profit} ]

The network operator wants:

[ \min \text{network stress} ]

Government wants:

[ \min \text{emissions} ]

These objectives coexist in the same system.

There is no single natural objective shared by everyone.

This is one reason markets exist:

to coordinate participants with different objectives.

Multi-objective optimisation versus multiple decision-makers

These ideas are related, but they are not the same.

A central optimiser might have several objectives:

[ f_1,f_2,f_3 ]

That is multi-objective optimisation.

But a system might instead contain several independent decision-makers:

[ A,B,C ]

each with their own objective:

[ U_A,U_B,U_C ]

Now we have a strategic interaction between participants.

That leads toward:

game theory
mechanism design
markets
distributed decision-making

We will return to these ideas later.

Fairness can be an objective

Suppose we allocate a scarce resource among:

[ n ]

people.

We might optimise:

[ \max \sum_i U_i ]

to maximise total utility.

But we might also care about inequality.

We could therefore introduce a fairness objective:

[ \min F(x) ]

where (F) measures some form of disparity.

Now the system might trade:

[ \text{efficiency} ]

against:

[ \text{fairness} ]

This makes an important tension explicit.

But fairness is not one thing

Different fairness objectives can produce different answers.

We might care about:

equal allocation
equal opportunity
equal outcomes
proportionality
need
contribution
cost causation
protecting the worst-off

So writing:

[ \max \text{fairness} ]

does not solve the problem.

We still need to define:

What do we mean by fair?

Reliability can also conflict with efficiency

Suppose System A costs:

[ £1\text{ billion} ]

and is reliable:

[ 99% ]

System B costs:

[ £1.5\text{ billion} ]

and is reliable:

[ 99.99% ]

Which should we choose?

The extra reliability may be extremely valuable.

Or it may be excessive.

The optimiser can show the trade-off.

It cannot decide society's tolerance for risk unless that preference is encoded.

Robustness adds another objective

We may also care about how a solution performs when assumptions are wrong.

Suppose Solution A is cheapest under the expected future.

Solution B costs slightly more but performs well across many possible futures.

Then we might care about:

[ \min \text{expected cost} ]

and:

[ \min \text{worst-case cost} ]

These are different objectives.

A system designed only for the expected case may be fragile.

A robust system may deliberately sacrifice some expected performance for resilience.

The cheapest solution is not always the best solution

This is worth stating clearly.

Suppose two bridges satisfy minimum safety requirements.

Bridge A:

[ £100\text{ million} ]

with expected lifetime:

[ 40\text{ years} ]

Bridge B:

[ £120\text{ million} ]

with expected lifetime:

[ 100\text{ years} ]

If we optimise:

[ \min \text{construction cost} ]

we choose A.

If we optimise:

[ \min \text{whole-life cost} ]

we may choose B.

If we also include:

resilience
maintenance disruption
environmental impact

the answer may change again.

What appears to be an optimisation problem is often first a problem-definition problem.

Beware of objective overload

There is also a danger in the opposite direction.

If we include every conceivable objective:

[ f_1,f_2,\ldots,f_{100} ]

the optimisation may become:

difficult to solve
difficult to explain
difficult to govern

Some objectives may also duplicate one another.

Good modelling requires discipline.

The goal is not:

include everything imaginable.

It is:

represent the things that materially determine whether the system is doing its job.

Objectives should come from purpose

A useful starting point is:

What is this system actually for?

If the purpose of a hospital is to improve health, its objectives should connect to health.

If the purpose of an electricity system is to provide reliable energy services, its objectives should connect to that purpose.

If the purpose of government is to serve citizens, its optimisation metrics should connect to outcomes experienced by citizens.

Metrics should follow purpose.

Purpose should not be reconstructed from whatever metrics happen to be easy to measure.

Stakeholders may disagree

Multi-objective optimisation becomes especially important when different groups value outcomes differently.

Suppose we are choosing between two infrastructure projects.

Residents may prioritise:

[ \text{quality of life} ]

Businesses may prioritise:

[ \text{transport connectivity} ]

Government may prioritise:

[ \text{economic growth} ]

Environmental groups may prioritise:

[ \text{ecological impact} ]

There may be no objective weighting everyone agrees with.

This is not necessarily a failure of optimisation.

It may simply reveal a genuine political disagreement.

Optimisation can expose disagreement rather than hide it

This is one of the most useful roles for multi-objective optimisation.

Instead of pretending there is one universally correct answer, we can show:

If you prioritise A, this happens.

If you prioritise B, this happens.

If you require C, these options disappear.

The model becomes a tool for understanding choices.

Not a machine for pretending value judgments do not exist.

Humans can remain in the loop

Imagine an optimiser generates ten Pareto-optimal infrastructure plans.

Each represents a different combination of:

cost
emissions
reliability
environmental impact

Humans can then examine those options.

This can be much more transparent than hiding all values inside one enormous objective function.

AI and optimisation can therefore support human decision-making without necessarily replacing it.

Interactive optimisation

We can go further.

A decision-maker might say:

Reliability below 99.9% is unacceptable.

The optimiser removes those solutions.

Then:

Show me solutions below £2 billion.

The feasible set shrinks again.

Then:

Among those, minimise emissions.

This creates an interactive process between:

human preferences
mathematical optimisation

The computer explores possibility.

The human helps determine desirability.

AI can help explore objective spaces

As optimisation problems become more complex, AI can help us:

identify candidate solutions
estimate outcomes
search large decision spaces
explain trade-offs
visualise Pareto frontiers
test alternative preferences

But the AI still needs to know:

what outcomes matter.

Prediction does not answer that question.

Prediction and multi-objective optimisation

Suppose an AI predicts:

[ P(\text{patient benefits from treatment})=0.82 ]

That prediction might feed into an objective such as:

[ \max \text{expected health benefit} ]

But the hospital may simultaneously want to:

[ \min \text{waiting time} ]

[ \min \text{cost} ]

and:

[ \max \text{fairness} ]

The prediction is therefore only one input into a much larger decision problem.

This is exactly why:

prediction is not decision-making.

A useful mathematical formulation

A general multi-objective problem might be written:

[ \min_{x\in\mathcal{F}} \begin{bmatrix} f_1(x) \ f_2(x) \ \vdots \ f_m(x) \end{bmatrix} ]

where:

[ \mathcal{F} ]

is the feasible set.

Unlike ordinary scalar optimisation, the output is not necessarily one obvious best solution.

Instead, we often seek the set of:

[ \text{non-dominated solutions} ]

forming the Pareto frontier.

From optimisation to choice

This gives us an important distinction:

Optimisation

What are the best achievable trade-offs?

Choice

Which trade-off do we actually prefer?

Those questions are related.

They are not identical.

An optimisation algorithm can often answer the first.

The second may require:

preferences
policy
ethics
democratic choice
market mechanisms
individual decisions

This distinction becomes increasingly important as AI moves into real-world decision systems.

A practical checklist

When facing a multi-objective problem, ask:

1. What are the objectives?

List them explicitly.

2. Do they conflict?

Can improving one make another worse?

3. Which outcomes are objectives and which are constraints?

Not everything should be tradeable.

4. Are the objectives measured in different units?

Do they need normalisation?

5. Are weights being used?

Who chose them?

6. Are any objectives merely proxies?

Do they really represent the underlying purpose?

7. Are some solutions dominated?

Remove obviously inferior choices.

8. What does the Pareto frontier look like?

What are the genuine best trade-offs?

9. Who should choose between those trade-offs?

An engineer?

A user?

A market?

A regulator?

A democratic process?

10. Can the decision be explained?

People should be able to understand what was gained and what was sacrificed.

Extending our framework

Our decision framework is becoming richer.

Prediction

What is likely to happen?

Objectives

What outcomes do we care about?

Utility and cost

How do we represent value?

Constraints

What is possible or permissible?

Optimisation

Which feasible solutions perform well?

Multi-objective optimisation

What are the best trade-offs when several outcomes matter?

Choice

Which trade-off do we actually want?

Decision

What do we do?

This is much closer to how important real-world systems actually work.

The key takeaway

Single-objective optimisation asks:

[ \boxed{\text{What is the best solution according to one objective?}} ]

Multi-objective optimisation asks:

[ \boxed{\text{What are the best available trade-offs between several objectives?}} ]

That difference is fundamental.

When objectives conflict, there may be no universally optimal solution.

Instead, we may have a Pareto frontier of good solutions:

improving one objective requires sacrificing another.

The optimiser can help us understand that frontier.

But it cannot make our values disappear.

Whenever someone presents an apparently optimal AI-generated decision, ask:

What objectives were included?

What was left out?

What was allowed to trade against what?

And who decided which trade-off was best?