Multi-objective

Optimization
Dr. Sobia Tariq Javed
Outline
• The objectives of this lecture are to:
• Multi-objective Optimization
• Address the design issues of evolutionary multi-objective optimization
algorithms
• Explore ways to handle Constraints
Multi-objective
Optimization
• Multiple Objectives: The problem has
several goals to optimize, often
conflicting (e.g., cost vs. quality).

• Trade-offs: Improving one objective may

degrade another, requiring compromises.

• Pareto Front: The set of all Pareto-

optimal solutions, showing the trade-offs
between objectives.
Example-
Car Buying
• Problem: Optimizing Car
Buying.
• Objectives:
• Minimize costs.
• Maximize comfort.
• Conflict: Increased
comfort often leads to
higher costs, requiring a
trade-off between
affordability and luxury.
Example-Supply
Chain
Management
• Problem: Optimizing supply
chain operations.
• Objectives:
• Minimize transportation
costs.
• Maximize delivery speed.
• Minimize environmental
impact (e.g., CO₂
emissions).
• Conflict: Faster deliveries often
result in higher costs and
emissions.
Example-
Environmental
Management
• Problem: Managing a water reservoir
system.
• Objectives:
• Maximize water storage for drought
conditions.
• Minimize flood risks by reducing
water levels.
• Ensure water availability for
agriculture and energy production.
• Conflict: High water levels for storage
may increase flood risk.
Single objective optimization
Objective function – Minimize f(x)
f(x)

X is better than
Multi-objective optimization
Objective function – Minimize f(x) = (f1(x), f2(x), f3(x))
f1(x) f2(x) f3(x)

is better than
Multi-objective optimization
Objective function – Minimize f(x) = (f1(x), f2(x), f3(x))
f1(x) f2(x) f3(x)

is ???? than
Multi-Objective -Dominance
Relationship
• A solution x dominates another solution y if
• x is no worse than y in all objectives And
• x is strictly better than y in at least one of the objectives.
Dominated
Solution
• A solution x is said to
dominate the other
solution y if
• the solution x is no • 1 Vs 2: ?
worse than y in all • 1 Vs 5:?
objectives and • 1 Vs 4:?
• the solution x is strictly
better than y in at least
one objective
Pareto Optimal Front
• Non-Dominated Solution Set
• The non-dominated solution set consists of all solutions within a given set that are not dominated
by any other solution in the set.

• Pareto Optimal Set

• The subset of the feasible decision space containing all non-dominated solutions is called the
Pareto optimal set.

• Pareto Optimal Front

• The boundary formed by mapping all points in the Pareto optimal set to the objective space is
referred to as the Pareto optimal front.
Pareto Dominance
Dominates x≺y Weakly dominates x≼y

Non-dominate
2D-case
f1(x) f2(x)
f1(x)

f2(x)
Dominance in 2D
f1(x)

dominated

dominating incomparable

f2(x)
Pareto Optimal Front
Min-min Min-max

Max-min Max-max
Pareto Optimal Solutions
• Pareto optimal solutions are the non-dominated solutions

• Most multi-objective optimization methods use this domination concept to search

for non-dominated solutions

• Solution 3 and 5 are non-dominated solutions 6

5 2

f2(minimize)
4
3 4
2 1 5
1 3
0
8 9 11 12 16

f1 (maximize)
 Decision and Objective Space

Eckart Zitzler, A Tutorial on Evolutionary Multiobjective Optimization. ETH Zurich, 2002.

Pareto Front
f1(x)

f2(x)
Pareto front = { }
Algorithm Design issues
Optimization Goal?
• Find all Pareto-optimal solutions?
• How should the decision maker handle 10000 solutions?
• Find a representative subset of Pareto set?
• Many problems are NP-hard
• What does representative really mean?
• Find a good approximation of the
Pareto set?
• What is a good approximation?
• How to formalize intuitive
understanding:
• Close to the Pareto-front
• Well distributed
Algorithm Design issues
Evolutionary Multiobjective Optimization (EMO) Principles
Two ideal goals of multi-objective optimization:

• Convergence
• Find a (finite) set of solutions which lie on the Pareto-
optimal front.
• Diversity
• Find a set of solutions which are diverse enough to
represent the entire range of the Pareto-optimal front.

Basic aim: Guiding the search towards the Pareto set and
keeping a diverse set of nondominated solutions.
 Algorithm Design issues
2. (Diversity Preservation)

3. (Elitism)

1. (Fitness assignment)

Eckart Zitzler, A Tutorial on Evolutionary Multiobjective Optimization. ETH Zurich, 2002.

1. Fitness Assignment Strategies

• Unlike single objective, multiple objectives exists.

• Fitness assignment and selection go hand in hand.
• Fitness assignment can be classified in to following categories:
• A) Aggregation-based
• e.g., MOEA/D
• B) Criterion-based (Objective-based)
• VEGA
• C) Dominance-based
• NSGA-II
A) Aggregation-based (scalarization
based)
• Aggregation Approach: Combines multiple objectives into a single
parameterized objective function to generate a trade-off surface.
• Systematic Variation: Parameters are adjusted during optimization to find a
set of nondominated solutions instead of just one.

• Weighted-Sum Aggregation: Some MOEAs use a weighted-sum approach, where

weights change dynamically during evolution.

• Pareto Dominance in Fitness Calculation: The concept of evaluating fitness

based on Pareto dominance was introduced by Goldberg.
A) Aggregation-based (scalarization
based)

Parameters -
weights

w1f1(x) + w2f2(x),…, wkfk

f1(x), f2(x),…, fk(x) F
Or, max(wi(fi - zi ))
A) Aggregation-based -Advantages
• Simple Implementation:
• Easy to integrate into traditional optimization frameworks.
• Reduces multi-objective optimization to a single-objective problem.
• Computationally Efficient:
• Faster convergence compared to Pareto-based methods.
• Requires fewer function evaluations.
• Allows Fine Control Over Trade-offs:
• Weight parameters can be adjusted to prioritize specific objectives.
• Useful when decision-makers have clear preferences.
A) Aggregation-based -Disadvantages
• Difficult to Define Appropriate Weights:
• Choosing the right weights can be subjective and problem-dependent.
• Small changes in weights may lead to very different solutions.
• Cannot Capture Non-Convex Pareto Fronts:
• Weighted-sum methods struggle with non-convex trade-off surfaces.
• Some Pareto-optimal solutions may be missed.
A) Aggregation-based -Disadvantages
• Loss of Diversity in Solutions:
• May favor certain objectives over others, leading to biased results.
• Difficult to obtain a well-distributed set of Pareto-optimal solutions.

• Lack of Adaptability:
• Fixed weights do not adapt dynamically during evolution.
• Requires multiple runs with different weights to explore the entire Pareto front.
B) Criterion-based (Objective-based)
• These methods switch between objectives during the selection phase.
• Different objectives may influence the selection of individuals at different times.
• Example: Vector Evaluated Genetic Algorithm (VEGA) by David Schaffer.
• Subpopulations are created, each evaluated based on a different objective.
• Ensures that solutions for all objectives are explored separately.
B) Criterion-based (Objective-based)
• Advantages:
• Simple & Easy to Implement (can extend single-objective algorithms).
• Finds Individual Best Solutions (useful when independent optimal solutions for each objective are
needed).

• Disadvantages:
• Each solution considers only one objective (ignores the trade-offs).
• May get stuck in local optima (reduces diversity in the final solution set).
B) Criterion-based -Example
• A company wants to design a lightweight yet strong material for an aircraft
component. The optimization problem has two conflicting objectives:
• 1. Minimize Weight (kg) → Reduce material usage.
• 2. Maximize Strength (MPa) → Ensure durability.

• Apply VEGA
• 1) Initialization
• A population of 10 solutions is randomly selected from the available materials.
• Each solution is a material selection for the component.

• 2) Subpopulation Creation
• VEGA splits the population into two equal subpopulations:
• Subpopulation 1 → Optimizes Weight
• Subpopulation 2 → Optimizes Strength
B) Criterion-based -Example
• 3. Selection Process
• In Subpopulation 1, solutions are ranked by minimum weight.
• In Subpopulation 2, solutions are ranked by maximum strength.
• The best-ranked solutions from each subpopulation are combined to form the next generation.

• 4. Crossover & Mutation

• Crossover: Materials can be swapped between solutions to create new combinations.
• Mutation: A randomly selected material can be replaced with another to explore more options.
• 5. Repeat for Multiple Generations
• The algorithm iterates, maintaining diversity and evolving toward a Pareto-optimal front, balancing weight vs.
strength.
C) Dominance based
• Dominance Metrics in Multi-Objective Optimization
• In Pareto-based evolutionary algorithms, dominance metrics help in ranking solutions based on how they
compare with others.
• 1) Dominance Rank (DR) → "How many individuals dominate a given solution?"
• Definition: The number of individuals in the population that dominate a given solution.
• Lower is better! A solution with DR = 0 is non-dominated (Pareto front 1).
• Example: If Solution A is dominated by three other solutions, then DR(A) = 3.

• 2) Dominance Count (DC) → "How many individuals does a solution dominate?"

• Definition: The number of individuals that a given solution dominates.
• Higher is better! A strong solution dominates many weaker solutions.
• Example: If Solution B dominates five other solutions, then DC(B) = 5.
C) Dominance based
• 3) Dominance Depth (DD) → "At which Pareto front is a solution located?"
• Definition: The layer number (Pareto front index) in which a solution lies.
• Front 1 contains non-dominated solutions.
• Front 2 contains solutions dominated only by those in Front 1, and so on.
• Example: If Solution C is in the second Pareto front, then DD(C) = 2.
C) Dominance based
• Example Population & Dominance Metrics
• Consider a multi-objective problem with three solutions (A, B, C) optimizing two conflicting objectives (e.g.,
cost vs. quality).

• Interpretation:
• Solution A is the best → It is non-dominated (Front 1, DR = 0, DC = 2).
• Solution B is dominated only by A → It is in Front 2.
• Solution C is the worst → It is dominated by both A and B (Front 3).
C) Dominance based
2. Diversity Preservation
• Ensuring a well-distributed Pareto front is crucial for obtaining a diverse set of trade-
off solutions. Diversity preservation prevents solutions from clustering in certain
regions while maintaining coverage across all objectives.

• The chance of an individual being selected depends on the density of solutions in its
vicinity:
• Increases: If the number of nearby solutions is low (promotes diversity).
• Decreases: If the number of nearby solutions is high (prevents overcrowding).

• Common Diversity Preservation Methods

• 1. Crowding Distance (NSGA-II)
• 2. Strength Pareto Ranking (SPEA2)
• 3. Reference-Point-Based Methods (MOEA/D)
3. Elitism-Environmental Selection
• Ensures good solutions are retained even if new generations produce worse
solutions.
• Reduces the risk of genetic drift, where random variations might eliminate optimal
solutions.
• Helps maintain diversity while guiding the population toward the Pareto front.
3. Elitism-Environmental Selection
• Selection Criteria for Environmental Selection
1. Dominance:
• Only non-dominated solutions (Pareto-optimal) are kept.
• Dominated solutions (worse in all objectives) are discarded.

2. Density-Based Selection:
• Solutions in less crowded regions are preferred.
• Ensures diversity by avoiding too many solutions in one part of the search space.
• Example: Crowding Distance in NSGA-II.
3. Elitism-Environmental Selection
• Selection Criteria for Environmental Selection
3. Time-Based Selection:

• Older, high-quality solutions are favored over newly introduced solutions.

• Prevents frequent replacement of good solutions.
• Example: Used in steady-state evolutionary algorithms.

4️. Chance-Based Selection:

• Each solution has an equal probability of being retained in the elite archive.
• Helps maintain exploration by not always favoring the best-known solutions.
Techniques to solve MOP
• 1) A Priori Methods
• Definition: Preferences are defined before the optimization process starts.
• How it Works: The decision-maker assigns weights, goals, or priorities to objectives in
advance.
• Example: Weighted Sum Method (assigning importance to each objective).
• Pros: Simple and fast.
• Cons: Requires accurate preference estimation upfront, may miss other valuable
solutions.
Techniques to solve MOP
• 2) A Posteriori Methods:
• Definition: Optimization is performed first, and preferences are applied after generating
a set of solutions.
• How it Works: The algorithm finds the Pareto front, and the decision-maker selects the
preferred solution from this set.
• Example: NSGA-II, MOPSO (evolutionary algorithms providing diverse solutions).
• Pros: Provides a range of trade-off solutions, flexible for decision-making.
• Cons: Computationally intensive, especially for large problems.
Techniques to solve MOP
• Interactive Methods:
• Definition: Preferences are refined during the optimization process in an iterative
manner.
• How it Works: The decision-maker gives feedback after each iteration to guide the
search toward more preferred regions.
• Example: Reference Point-Based Methods, Trade-off Analysis.
• Pros: Dynamic adjustment to evolving preferences, better alignment with decision-
maker’s goals.
• Cons: Requires continuous involvement, can be time-consuming.
Techniques to solve MOP
 Optimization and Decision Making

Eckart Zitzler, A Tutorial on Evolutionary Multiobjective Optimization. ETH Zurich, 2002.

Why EC algorithms?
• Evolutionary algorithms (EAs) are well-suited for solving multi-
objective optimization problems because they work with a
population of solutions simultaneously. This enables the
discovery of multiple Pareto-optimal solutions in a single run,
unlike traditional mathematical programming techniques that
require multiple runs.
• Additionally, EAs are robust and less affected by the shape or
continuity of the Pareto front, making them effective for complex,
non-linear problems.
References

• Multi-Objective Optimization using Artificial Intelligence Techniques

Chapter 2
• A.E. Eiben and J.E. Smith Chapter 12