ACS6124 Multisensor and Decision Systems

Part II – Decision Systems for Engineering Design

Lecture 4: Multi-objective Optimization II

Prof. Robin Purshouse

University of Sheffield

Spring Semester 2019/2020

1 Lecture overview
This lecture continues the introduction to multi-objective optimization as a core ele-
ment of a decision system for engineering design.
In this lecture, we will cover:
• Decomposition-based algorithms;
• Set-based algorithms.
During the lecture we will continue to work with the unconstrained multi-objective
optimization problem (MOP) from the previous lecture:

minimise z = f (x, p)
x
x∈D
p←P (1)
where x is a candidate design from a design space D, z is a vector of perfor-
mance criteria, and f (.) is a vector of evaluation functions that estimate how a design
will perform against the criteria when using a nominal parameterisation p = p0 from
parameter space P.

2 Decomposition-based decision systems

2.1 Introduction to decomposition
Decomposition-based methods transform the vector of performance criteria z into a
scalar value ζ. This is achieved using a scalarising function ζ = s(z, w, zref ) where w
is a vector of weighting parameters and zref is a reference point in performance space.
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

The scalarising function s(.) is chosen in such a way that optimization of the func-
tion will yield a Pareto optimal design from the efficient set X? (see Equation 4 from
Lecture 3). For the purposes of a posteriori decision making, the idea is that, by
changing w and/or z? and then re-optimizing s(.), designs from different locations on
the Pareto front can be found.

2.2 Scalarising functions

In this section, we consider the most popular forms of scalarising function used in
decomposition-based approaches.

2.2.1 Weighted sum

The weighted sum scalarising function is given by:
M
X
ζ = s(x|w) = wi zi (x) (2)
i=1
PM
where wi ≥ 0 ∀i, i=1 wi = 1, and M is the number of objectives.
An example of the weighted sum scalarising function for a given w is shown in
Figure 1. For convex Pareto fronts all points on the front can be found, given a suitable
choice of w. However, concave regions of Pareto fronts cannot be found using
this approach. The reason is that the boundaries of the concave region will always
produce a better value for ζ than the interior of the concave region for any choice of
w. We must therefore be careful to ensure that we are dealing with a convex Pareto
front prior to application of any weighted sum decomposition method.

2.2.2 Reference point

The reference point or achievement scalarising function is given by:
M
X
ζ = s(x|zref ) = (zi (x) − ziref ) (3)
i=1

where ziref is the ith component of a reference point in performance space.

If zref relates to a level of achievement that is of interest to the decision maker
then optimization of the reference point scalarisation function will produce the closest
Pareto optimal point to that level of achievement. Further, all points on the Pareto front
can be found by specifying alternative choices for zref .

Part II – Decision Systems for Engineering Design 2

Lecture 4: Multi-objective Optimization II
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

z2

z1

Figure 1: Weighted sum scalarisation of eight performance vectors for an arbitrary

weight vector w1 . The arrows indicate the mapping to the line ζ(x|w = w1 ). The
dashed lines indicate contours in ζ.

Part II – Decision Systems for Engineering Design 3

Lecture 4: Multi-objective Optimization II
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

2.2.3 Weighted metric

The weighted metric combines and generalises the weighted sum and reference point
approaches. The scalarising function is given by:

M
!1/l
X
ref
ζ = s(x|w, z , l) = wi (zi (x) − ziref )l (4)
i=1

where 1 ≤ l ≤ ∞ is the choice of norm and other variables are as previously

introduced.
For suitably varying w and/or zref , optimization of the corresponding weighted met-
ric scalarising function will yield the full range of solutions on the Pareto front. However
note that with l = 1 and zref fixed, the function is equivalent to the weighted sum and
will not be able to identify concave areas of the front.

2.2.4 Tchebycheff metric

A popular configuration of the weighted metric scalarisation function given in Equa-
tion 4 is to set l = ∞ and zref to the ideal point of best possible performance in every
criterion: zref = z? . The resulting function is known as the Tchebycheff scalarising
function and can be equivalently formulated as:

ζ = s(x|w, zref = z? , l = ∞) = max [wi (zi (x) − zi? )] (5)

i

An example of the Tchebycheff scalarising function for a given w is shown in Fig-

ure 2.
Minimisation of Equation 5 will yield at least a weakly Pareto optimal solution. As
usual, by varying w, multiple such solutions can be found. To ensure that the search
progresses beyond weakly optimal solutions to the Pareto optimal solutions (i.e. down
or across the lines of equal ζ according to the (l = ∞)-norm), Equation 5 can be
extended to form the augmented Tchebycheff scalarising function:

M
X
ζ = s(x|w, zref = z? , l = ∞, ρ) = max [wi (zi (x) − zi? )] + ρ (zi (x) − zi? ) (6)
i
i=1

where ρ is a sufficiently small positive scalar constant.

2.3 Normalisation
Decomposition-based approaches require objectives to be aggregated together. Such
aggregation can be problematic in cases where the different performance criteria use
different units (e.g. grams of CO2 for a fuel economy criterion; kilometres per hour for
a maximum speed criterion) or are of different orders of magnitude.

Part II – Decision Systems for Engineering Design 4

Lecture 4: Multi-objective Optimization II
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

z2

z1

Figure 2: Tchebycheff scalarisation of eight performance vectors for the given weight
vector w1 . The arrows indicate the mapping to the line ζ(x|w, zref = 0, l = ∞). The
dashed lines indicate contours in ζ.

Part II – Decision Systems for Engineering Design 5

Lecture 4: Multi-objective Optimization II
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

nadir
vector znad

z2

ideal
vector z*
z1

Figure 3: Ideal and nadir vectors of performance, with respect to the Pareto front

To avoid problems, it is recommended to normalise the objectives prior to scalari-

sation. Normalisation is usually performed according to the following equation:

zi (x) − zi?
z̄i (x) = (7)
zinad − zi?
where zi? is the ith component of the ideal vector (introduced previously) and zinad is
the ith component of the nadir vector of worst performance across all Pareto optimal
solutions. Figure 3 illustrates ideal and nadir vectors for a two-dimensional perfor-
mance space. Note that estimating the ideal and nadir vectors in advance of opti-
mization can be problematic (particularly the estimation of nadir vectors) – this is a
key limitation of the decomposition-based approach.

2.4 MOEA/D
The multi-objective evolutionary algorithm based on decomposition (MOEA/D) is the
most well-known optimizer based on the decomposition concept. This section of the
lecture provides an introduction to the key concepts that underpin the algorithm.

Part II – Decision Systems for Engineering Design 6

Lecture 4: Multi-objective Optimization II
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

w1 w2 w3
0 0 1
0 0.333 0.667
0 0.667 0.333
0 1 0
0.333 0 0.667
0.333 0.333 0.333
0.333 0.667 0
0.667 0 0.333
0.667 0.333 0
1 0 0

Table 1: An example weight vector set W for a three-objective problem with three
divisions per objective

2.4.1 Scalarising functions as sub-problems

To use MOEA/D within a decision system it is first necessary to define a scalarisa-
tion function and associated parameters. The algorithm can work with any of the
scalarising functions introduced above but is usually configured to use the augmented
Tchebycheff. The scalarising function is then used to define a number of sub-problems
which will be optimized simultaneously – each sub-problem corresponds to a particu-
lar choice of weight vectors w.
The set of weight vectors should aim to capture the overall characteristics of the
Pareto front, typically aiming for an even coverage. In the basic MOEA/D algorithm,
weight vectors are constructed according to an {M, h}-simplex-lattice design:
( M   )
X 0 1 h
W = w = [w1 , . . . , wM ] | wi = 1 ∧ wi ∈ , ,..., ∀i (8)
i=1
h h h
where h defines the number of divisions for each objective. Under the simplex-
lattice design, the set of weight vectors constitutes all possible combinations of weights
that satisfy Equation 8. The total number of weight vectors produced will be N =
h+M −1
CM −1 . For example, a {3, 3}-simplex-lattice creates the set of 5 C2 = 10 weight
vectors shown in Table 1.
Note that the simplex-lattice design produces vectors that are evenly distributed in
the space of weights; but this will only produce an even distribution on the Pareto front
if the latter is linear and has been normalised correctly. Therefore for Pareto fronts
that are convex, concave, or of mixed geometry, or cases where normalisation is not
perfect, it may be better to use a different approach. This is an advanced topic that
lies outside the scope of the module.

2.4.2 Sub-problem neighbourhoods

Once the sub-problems have been defined, a neighbourhood B(j) around each sub-
problem j is established. The neighbourhoods are defined by calculating the Eu-
clidean distances between all pairs of weight vectors and then, for every sub-problem

Part II – Decision Systems for Engineering Design 7

Lecture 4: Multi-objective Optimization II
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

j, choosing the sub-problems with the closest T weight vectors as the neighbours of
j. Here the sub-problem itself is included in the count of T , so a neighbourhood of
size T = 2 would include sub-problem j and one other sub-problem that is closest to
j in weight space.

2.4.3 The algorithm

MOEA/D allocates a design to each sub-problem. It then uses the neighbourhoods
around the sub-problems to perform localised selection-for-variation, variation, and
selection-for-survival operations.
1. Initialise the iteration counter t = 0.
2. Generate a candidate design xtj for every sub-problem, i.e. j = 1 . . . N .
3. Estimate the ideal vector ẑ? using the initial population or prior knowledge.
4. For every sub-problem j = 1 . . . N :
(a) Randomly select the current designs from two sub-problems in the neigh-
bourhood B(j) and apply variation operators to these designs to generate
a new candidate design yjt .
(b) Evaluate new design yjt .
(c) Re-estimate the ideal vector: ∀i ∈ 1, . . . , M , if zi (yjt ) < ẑi? then ẑi? = zi (yjt )
(d) Across the neighbourhood of j, replace current design xtk with design yjt if
s(z(yjt ), wk , ẑ? ) ≤ s(z(xtk ), wk , ẑ? ) for neighbour k.
5. Increment the iteration counter t = t + 1.
6. If a convergence criterion is satisfied then stop; otherwise go to Step 4.
MOEA/D has three principal benefits as an algorithm: (i) it is parallelisable; (ii) it
allows specific directions to be targeted (assuming weight vectors can be defined ap-
propriately); (iii) the localisation of its operators helps improve convergence efficiency
by only sharing information across similar areas of performance space – this latter
feature helps avoid the phenomenon of lethals, where recombination operations on
very different parent solutions tend to produce poor performing child solutions.
MOEA/D’s principal drawback relates to the need to perform effective normalisa-
tion. The performance of the algorithm may also be sensitive to the choice of its
neighbourhood size parameter, T .

3 Set-based decision systems

3.1 Background
So far we have considered the Pareto-based and decomposition-based classes of
multi-objective optimizers. Both these classes are population-based, but are con-
cerned with the performance of individual candidate designs in comparison to others.

Part II – Decision Systems for Engineering Design 8

Lecture 4: Multi-objective Optimization II
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

It is hoped that good individual levels of performance, when taken across a popula-
tion, will yield a good overall approximation set for the Pareto front. The final class
of approach we will consider are set-based methods, which explicitly consider the
performance of a population (or set) of solutions as a whole.
Unusually, set-based methods arose from performance testing of existing multi-
objective optimizers. Performance indicators were developed that summarised the
approximation set that was generated by one optimizer and allowed it to be compared
to the approximation set generated by a different optimizer. Researchers quickly re-
alised that these performance measures could actually be used as fitness measures
to drive the convergence of an algorithm. For this reason, set-based methods are also
sometimes described as indicator-based methods.
A number of different performance indicators have been harnessed for use in set-
based optimization. In this module, we will focus on one particular indicator known as
hypervolume or the S-metric.

3.2 Hypervolume
Given an approximation set, A, the hypervolume measures the volume of the perfor-
mance space that is weakly dominated by the approximation set’s members, where
the volume is bounded by some dominated reference point, zref . Formally, the hyper-
volume, S(A, zref ), is defined as:
 
[ n o
S(A, zref ) = µ  z0 : z(x)  z0  zref  (9)
z(x)∈A

where µ(.) is the Lebesgue measure of the set (or, in more simple terms, the
volume defined by the set).
As an indicator, hypervolume has two very desirable properties: (i) if S(A1 , zref ) >
S(A2 , zref ) then approximation set A1 cannot be dominated by approximation set A2
(i.e. it cannot be worse in terms of Pareto dominance); and (ii) hypervolume is max-
imised, for a given zref , if and only if A is the set of all Pareto optimal points. An
example calculation of hypervolume for a performance space with two objectives is
shown in Figure 4.

3.3 SMS-EMOA
The S-metric selection evolutionary multi-objective optimization algorithm (SMS-EMOA)
is one of the most well-known set-based optimizers. The algorithm attempts to find a
set of Pareto optimal designs that maximises the hypervolume indicator, subject to a
defined finite approximation set size.
Like NSGA-II, SMS-EMOA uses non-dominated sorting as the primary compo-
nent of its selection mechanism. However, the crowding distance used in NSGA-II is
replaced with a measure based on hypervolume.

Part II – Decision Systems for Engineering Design 9

Lecture 4: Multi-objective Optimization II
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

zref

z(x1)

z2
z(x2)

z(x3)

z(x4)
A={z(x1), z(x2), z(x3), z(x4), z(x5)} z(x5)

z1

Figure 4: Illustration of the hypervolume indicator S(A, zref ) for an approximation set
with five members. The hypervolume is shown by the grey area.

Part II – Decision Systems for Engineering Design 10

Lecture 4: Multi-objective Optimization II
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

3.3.1 The algorithm

SMS-EMOA proceeds as follows:

1. Initialise the iteration counter t = 0.

2. Generate a population P[0] of N candidate designs.
3. Randomly select two designs from the population P[t] and apply variation oper-
ators to these designs to generate a new candidate design yt .
4. Evaluate new design yt .
5. Apply the Reduce procedure to generate the new population P[t + 1] from {P[t] ∪
yt }.

6. Increment the iteration counter t = t + 1.

7. If a convergence criterion is satisfied then stop; otherwise go to Step 3.

3.3.2 The Reduce procedure

The Reduce procedure operates in a similar fashion to NSGA-II’s selection-for-survival
operator. The new population is defined as the null set, P[t + 1] = ∅, and is incre-
mentally filled using ranks obtained from the non-dominated sorting procedure (see
Section 4.2.1 in Lecture 3). When a rank is reached that would lead to an overspill
of the population, the designs in the final rank are inspected for the contributions that
each makes to the overall hypervolume associated with that rank. The design with
the lowest hypervolume contribution is then discarded. An example of this process is
shown in Figure 5.
An advantage of the SMS-EMOA algorithm is that the ranking of solutions ac-
cording to hypervolume contribution is invariant to scaling of the objectives, making
normalisation less of a concern in comparison to MOEA/D. However the hypervolume
calculation is sensitive to choice of the reference point zref . The hypervolume is also
computationally expensive to calculate analytically – although it can be approximated
more cheaply using Monte Carlo approaches.

4 Further reading
• Zhang Q., Li H. MOEA/D: A multi-objective evolutionary algorithm based on
decomposition. IEEE Transactions on Evolutionary Computation 2007;11:712–
731.
• Beume N., Naujoks B., Emmerich M., SMS-EMOA: Multiobjective selection based
on dominated hypervolume. European Journal of Operational Research 2007;181:1653–
1669.

Part II – Decision Systems for Engineering Design 11

Lecture 4: Multi-objective Optimization II
Prof. Robin Purshouse ACS6124 Multisensor and Decision Systems University of Sheffield

zref

z(x1)

z2
z(x2)

z(x3) z(y)

z(x4)
z(x5)

z1

Figure 5: Hypervolume contributions – a new candidate design y is considered for

inclusion in the new population of five solutions. The hypervolume contributions of
each design are shown by the grey areas. The new design would not be retained by
SMS-EMOA’s Reduce algorithm.

Part II – Decision Systems for Engineering Design 12

Lecture 4: Multi-objective Optimization II