University of Zagreb, Croatia

Lecture 5
Heuristic Optimization Methods:
Local Search, GRASP
Slides prepared by Nina Skorin-Kapov

Academic year 2020/2021

Graduate Studies Programme
 Outline
University of Zagreb, Croatia

 Basic Local Search elements:

 Neighborhood
 Initial
solution
 Neighbor selection

 Classical Local Search (LS)

 LS improvements
 Multi-start LS
…

 GRASP

Zavod za telekomunikacije 2
 Optimization methods
University of Zagreb, Croatia

Exact Approximate
A*
methods methods
Simplex- Constraint
based for LP programming Approximation
algorithms
Heuristic
Branch and Bound, Dynamic
Cut , Price (B&B, programming algorithms
B&C, B&P)

Improvement (meta)heuristics
Constructive heuristics
• Local search
• Greedy algorithms
•Nature inspired
• Tabu search, Genetic
Hybrid methods
algorithms, Simmulated
• GRASP Annealing, Ant colony
•Problem specific
heuristics

Zavod za telekomunikacije 3
 Basic Local Search elements
University of Zagreb, Croatia

 Main idea: iteratively improve the current

solution by applying local transformations to it
 Local Search is a generic heuristic approach
often used within more complex single-solution
based metaheuristics (Tabu search, GRASP,
Simulated Annealing,…)
 Basic elements:
 Neighborhood
 Initial
solution
 Neighbor selection

Zavod za telekomunikacije 4
 General Local search procedure
University of Zagreb, Croatia

Start Local Search

Generate an initial solution (the current
solution)
do
Generate a set of candidate neighboring
solutions from the current solution
Choose a better neighbor to become the new
current solution if it exists. If not,
stop.
While a better solution was found
End

Zavod za telekomunikacije 5
 Neighborhood
University of Zagreb, Croatia

 The neighborhood N(s) of a solution s in S is set of

solutions N (s)  S , where every solution s' N (s) can
be obtained from s by applying some elementary
transformation/perturbation, called a move, to s

 A neighborhood function N is a mapping N : S → 2S

 s' N (s) is called a neighbor of s

N(s)
s
S
Zavod za telekomunikacije 6
 Continuous neighborhood
University of Zagreb, Croatia

The neighborhood
of a continuous
N(s) problem in 2
s
dimensions

 In continuous optimization, neighborhood N(s)

in a continuous space is the ball with center s
and radius equal to ε with ε > 0.

Zavod za telekomunikacije 7
 Discrete neighborhood
University of Zagreb, Croatia

101
111
3 binary variables
100
110
Neighbors are
adjacent nodes in
the hypercube
001 011 representing all
solutions

000 010

 In discrete optimization, neighborhood N(s) is

represented by the set {s′/d(s′, s) ≤ ε}, where d
represents a given distance that is related to the move
operator.
We will consider mostly discrete optimization
Zavod za telekomunikacije 8
 Example: SAT
University of Zagreb, Croatia

(A ˅ B ˅ C) ˄(C ˅D) ˄ (E ˅A)…

 Representation: vector of n bits (n=number of
variables) A B C D E
1=TRUE
1 0 1 1 0 0=FALSE
 Neighborhood: “1 flip”
 Change the value (truth assignment) of 1 variable

0 0 1 1 0
1 1 1 1 0 n neighbors
1 0 0 1 0
1 0 1 0 0
1 0 1 1 1
Zavod za telekomunikacije 9
 Example: TSP
University of Zagreb, Croatia

 Representation: permutation of cities

Current solution s: A→B→C→D

1 2 4 Fitness:
Evaluation: 2
A B C D A F(s)=9

A 1 B

4 4 3 2

D C
2
Zavod za telekomunikacije 10
 Example: TSP
University of Zagreb, Croatia

 Neighborhood: “pairwise exchange”

Current 1
A→B→C→D A B
solution s

4 4 3 2
B→A→C→D
C→B→A→D D C
2
Neighborhood D→B→C→A n(n − 1)
N(s)
neighbors
A→C→B→D 2
n is the number of cities
A→D→C→B
4(4 − 1)
A→B→D→C = 6 neighbors
2
Zavod za telekomunikacije 11
 TSP: Evaluation of neighborhood
University of Zagreb, Croatia
Current solution s:: Fitness:
1 2 2 4 F(s)=9
A B C D A
Neighborhood N(s):
1 3 2 4 F(s)=10
B A C D B
2 1 4 2 F(s)=9
C B A D C
4 2 3 4 F(s)=13
D B C A D
3 2 4 4 F(s)=13
A C B D A
4 2 2 1 F(s)=9
A D C B A
1 4 2 3 F(s)=10
A B D C A
Zavod za telekomunikacije 12
 Neighborhood size
University of Zagreb, Croatia

 Design issue: size

 Often a trade-off between neighborhood

size/quality and the computational
complexity to explore it.
 Designing large neighborhoods can improve
the quality of the obtained solutions since
more neighbors are considered at each
iteration, but require more computational
time
Zavod za telekomunikacije 13
 Initial solution
University of Zagreb, Croatia

 2 main strategies:
 Random – fast, but may take longer to
converge
 Greedy – takes somewhat longer to
generate, but often leads to better quality
solutions (local optima)

 NOTE: using initial solutions with better

fitness values will not necessarily always
lead to better local optima.
Zavod za telekomunikacije 14
 Neighbor selection techniques
University of Zagreb, Croatia

 Best improving neighbor (descent method/hill climbing)

 Exhaustive exploration of the neighborhood, best neighboring
solution chosen.
 Can be time-consuming for large neighborhoods
 First improving neighbor
 Choose the first neighbor which improves the current solution;
partial (cyclic) evaluation of neighborhood (if a better solution
exists; otherwise exhaustive)
 Random improving neighbor
 One of the solutions which improves the current solution chosen at
random, partial or exhaustive exploration of neighborhood
 Best/random improving in reduced neighborhood
 Partial neighborhood exploration (reduced neighborhood can be
random or according to some strategy), best/random improving
neighbor of those explored is chosen
Zavod za telekomunikacije 15
 Neighborhood selection
University of Zagreb, Croatia

Neighbor
selection
Neighborhood N(s) f(s)
Initial Best improving:
solution s f(s) 0 0 1 1 0 12 f(s’)=14
1 1 1 1 0 10 First
1 0 1 1 0 10 1 0 0 1 0 14 improving:
1 0 1 0 0 9 f(s’)=12
1 0 1 1 1 11
Random
Reduced improving:
neighborhood: Best: f(s’)=14 f(s’)=12, 14 or 11
or random: f(s’)=14 or 11
Zavod za telekomunikacije 16
 Classical LS: steepest descent method
University of Zagreb, Croatia

Start LocalSearch
Generateinitialsolutions 0
s  s0
do
GenerateN( s)
Find s ' N ( s ) where f ( s ' )  f ( s ' ' ) , s'' N (s)
If f ( s ' )  f ( s )
s  s'
Else stop
while s was updated
return s
End
Zavod za telekomunikacije 17
 Local optimum
University of Zagreb, Croatia

 Always stops at the first local optimum

 Example: minimization

This is not
Obtained LS necessarily
solution: the first globally
local optimum optimal!
Zavod za telekomunikacije 18
 Impact of size of neighborhood
University of Zagreb, Croatia

Initial solution
Small
Large neighborhood
neighborhood

Obtained local
Obtained local optimum with
optimum with larger small
neighborhood neighborhood

Zavod za telekomunikacije 19
 Example
University of Zagreb, Croatia

 Objective and fitness function:

3 2
max f ( x) = x − 60 x + 900 x + 50
x = [0,31], integer
 Representation: x is shown in binary form with
5 bits
 Initial solution:

1 0 0 1 1 f(s)=2349

Zavod za telekomunikacije 20
 Example: neighborhood
University of Zagreb, Croatia

 Current solution:
s 1 0 0 1 1 f(s)=2349

 Neighborhood “1 flip”

ITERATION 1:
f(s)
0 0 0 1 1 2237 Best improving:
1 1 0 1 1 293 f(s’)=2923
N(s) New current
1 0 1 1 1 1177
1 0 0 0 1 2923 solution
1 0 0 1 0 2642

Zavod za telekomunikacije 21
 Example: neighborhood
University of Zagreb, Croatia

ITERATION 1: f(s) Best improving: f(s’)=2923

0 0 0 1 1 2237 New current solution: 4th bit
1 1 0 1 1 293 flipped
N(s) 1 0 1 1 1 1177
1 0 0 0 1 2923 s 1 0 0 0 1 f(s)=2923
1 0 0 1 0 2642
Generate neighborhood
ITERATION 2: f(s)
0 0 0 0 1 891
1 1 0 0 1 675 Best improving:
N(s) f(s’)=3186
1 0 1 0 1 1751
1 0 0 1 1 2349 New current
1 0 0 0 0 3186 solution

Zavod za telekomunikacije 22
 Example: neighborhood
University of Zagreb, Croatia

ITERATION 2: f(s) Best improving: f(s’)=3186

0 0 0 0 1 891 New current solution: 5th bit
1 1 0 0 1 675 flipped
N(s) 1 0 1 0 1 1751
1 0 0 1 1 2349 s 1 0 0 0 0 f(s)= 3186
1 0 0 0 0 3186
Generate neighborhood
ITERATION 3: f(s)
No better neighbor found!
0 0 0 0 0 50 Local search ends
N(s) 1 1 0 0 0 914 Final solution (local optimum):
1 0 1 0 0 2050
1 0 0 1 0 2642 1 0 0 0 0 f(s)= 3186
1 0 0 0 1 2923
(Global opt: f(s)=4050)
Zavod za telekomunikacije 23
 Local Search improvements
University of Zagreb, Croatia

Source: Talbi, “Metaheuristics: From Design to Implementation”, 2009

Zavod za telekomunikacije 24
 Multistart local search
University of Zagreb, Croatia

 Example: minimization, 3 initializations

2nd initial solution 1st initial solution

3rd initial solution

END

Already
found!
Still not
1st found local
2nd found local The best seen necessarily
solution is the optimum globally
optimum
final one optimal…
Zavod za telekomunikacije 25
 Outline
University of Zagreb, Croatia

 Basic Local Search elements:

 Neighborhood
 Initial
solution
 Neighbor selection

 Classical Local Search (LS)

 LS improvements
 Multi-start LS
…

 GRASP

Zavod za telekomunikacije 26
2
7
GRASP
University of Zagreb, Croatia

 GRASP: Greedy Randomized Adaptive Search

Procedure
 Multi-start or iterative single-solution
metaheuristic
 Each GRASP iteration is independent (no
search memory, except for incumbent solution)
 Each GRASP iteration consists of 2 phases:
1. Construction phase
2. Local Search phase

Zavod za telekomunikacije 27
2
8
GRASP Pseudo-code
University of Zagreb, Croatia

Start GRASP
Input : MaxIterations, seed
for k = 1,...MaxIterations
s  Constructi on _ Phase ( seed )
s´ Local _ Search _ Phase ( s )
Update solution ( s best , s´);
end
return s best ;
End GRASP

Zavod za telekomunikacije 28
2
9
GRASP: Construction Phase
University of Zagreb, Croatia

 The construction phase builds a feasible solution using a

randomized greedy algorithm
 If the obtained solution is not feasible, some repair procedure
should be applied, or a new solution created

 A randomized greedy algorithm uses the same principle

as regular greedy algorithms, but includes some
randomization to build different solutions in each run
 When building the solution, instead of selecting the element
which contributes most to the objective function (e.g., max profit,
min cost,…), one of the “best” elements (but not necessarily the
best) is randomly chosen

Zavod za telekomunikacije 29
3
0
GRASP: Construction Phase
University of Zagreb, Croatia

The randomized greedy algorithm

 In each iteration, the set of all elements that can
potentially be included in the partial solution are
ordered according to a greedy evaluation function
(e.g. minimizing cost)
 Then a subset of these elements is generated, called
the Restricted Candidate List (RCL), which contains
only the “best” elements
 The element to be included in the partial solution is
chosen randomly from the RCL using a uniform
probability distribution
 Then the partial solution is updated for the next
iteration

Zavod za telekomunikacije 30
3
1
GRASP Construction Phase Pseudocode
University of Zagreb, Croatia

Start Construction_Phase (seed)

s ← 0;
Evaluate candidate elements according to greedy
function;
while solution is not a complete solution do
Build RCL;
Select element e from RCL at random;
𝑠 = 𝑠 ∪ {𝑒};
Re-evaluate candidate elements using greedy function;
end
return 𝑠;
End Construction_Phase

Zavod za telekomunikacije 31
3
2
Restricted Candidate List (RCL)
University of Zagreb, Croatia
 The size of the RCL can be limited either by the number of elements
(cardinality-based) or by their quality with respect to the best candidate
element (value-based)
 Cardinality-based criteria
 The RCL is composed of the n best elements in terms of incremental cost
(i.e. the greedy evaluation function)
 Value-based criteria (more commonly used)
 All candidate elements (eϵC) whose incremental cost (c(e)) is better than a
threshold value enter the RCL
 The threshold is calculated as (assuming a minimization problem and
greedy function)
c min +  (c max − c min ), where c min = min{c(e) | e  C}, c max = max{c(e) | e  C}
 Solutions that enter the RCL are:

RCL  {e  C | c(e)  c min +  (c max − c min )}

 Parameter αϵ[0,1] can be tuned to desired value:
 If α=0 →pure greedy; If α=1 →pure random
Zavod za telekomunikacije 32
3
3
Local Search Phase
University of Zagreb, Croatia

 Since the solution obtained by the construction phase

may not be locally optimal, a local search procedure is
applied
 Traditionally simple local search is applied, but other
single-solution metaheuristics could also be used (e.g.
tabu search, simulated annealing,..)

Start Local_Search_Phase (s)

while Solution is not locally optimal (i.e., a better solution exists in its
neighborhood)
Find s´ ∈ N(s) with c(s´) < c(s)
𝑠 ← 𝑠´
end
return 𝑠;
End Local_Search_Phase

Zavod za telekomunikacije 33
 Case study: DCMR problem
University of Zagreb, Croatia

Delay-constrained Multicast Routing (DCMR)

 The rising development of real-time multimedia
applications
 Video/tele-conferencing, VoD, distance education,
distibuted games, etc.
 QoS constraints
 E2E delay, interreceiver delay jitter, packet loss
probability, min bandwidth
 Delay constrained multicasting routing (DCMR)
 Goal: finding the lowest cost multicast tree satisfying
a maximum delay constraint between the source and
destination nodes
Zavod za telekomunikacije 34
Recall the Minimum Steiner Tree Problem
University of Zagreb, Croatia

• Steiner tree: Given a graph G=(V,E) and a subset of

nodes D in V, a Steiner tree is a connected and acyclic
subgraph of G which spans all nodes in D
• A minimum Steiner tree is such a subgraph of minimum
weight in a weighted graph

(3) e
b
(4)
(3) (1) (3)
h
a d (2) f

(1) (2)
(4) (1) (2)
c
(4)
g

Zavod za telekomunikacije 35
Recall the Minimum Steiner Tree Problem
University of Zagreb, Croatia

Representation: set of Steiner nodes S represented by a

binary vector of size |V/D|
Decoder function: find an MST (Minimum Spanning Tree) of
sub-graph G composed only of nodes S and D
Fitness function: size of the found MST Add max delay
constraint!
Non-terminal nodes: [a d e f g]
(3) e
Representation: b
s = [0 1 0 1 1] (3)
(4)
1: included in solution (3) (1)
0: removed from graph h
a d (2) f
Solution:
the MST shown (1) (2)
(4) (1) (2)
Fitness function: c
(4)
f(s)=8 g

Zavod za telekomunikacije 36
 Example
(3, 4) University of Zagreb, Croatia

source a (2, 1) b (2, 1) c (5, 2) d

(2, 2)
e (3, 2) (1, 3)(3, 4)
(4,2)
(3, 1) (3, 1) h i j
Delay (2, 1) g (1, 4)
constraint (5, 2) (2, 1)
f (3, 1)
Δ=6 l
(3, 2) m (2, 3)
(1, 4) k
(x,y) →
x indicates cost, (1, 1)
(1, 3)
n
y indicates delay (3, 2) (2, 5)
o
p (2, 1) (1, 1)
s
(3, 3)
q (2, 1) (4, 3)
(4, 1) (6, 1)
r
u (4, 2)
Zavod za telekomunikacije t 37
 Problem reduction
(3, 4) University of Zagreb, Croatia

source a (2, 1) b (2, 1) c (5, 2) d

(2, 2)
(2, 2)
e (3, 2) (1, 3)(3, 4)
(4,2)
(3, 1) (3, 1) h i j
Delay (2, 1) g (1, 4)
constraint (5, 2) (2, 1)
f (3, 1) l
Δ=6
(3, 2) m (2, 3)
(1, 4) k
(1, 1)
(1, 3)
n
(3, 2) (2, 5)
o
p (2, 1) (1, 1)
s
(3, 3)
q (2, 1) (4, 3)
(4, 1) (6, 1)
r
u (4, 2)
Zavod za telekomunikacije t 38
 Problem reduction
University of Zagreb, Croatia

x x x

(3, 2)

(3, 3) y (3, 3) (5, 4)' (3, 3)

(2, 2)
z
z z
x d. x
(2, 2)

(3, 5) y (3, 5) (4, 3)'

(2, 1)

z z
Zavod za telekomunikacije 39
 The reduced graph
(3, 4) University of Zagreb, Croatia

a (4, 2)' c

(2, 1)
(3, 1) (4,2)
k (4, 5)'
(3, 1)
f (5, 2) g (1, 3)

(1, 4) (3, 2) o
(2, 5)
n
(1, 1)

(2, 1)
s
(4, 3)
(3, 3)
q
(4, 1) (6, 1) (2, 1)
r
u (4, 2)
Zavod za telekomunikacije t 40
 The reduced graph
(3, 4) University of Zagreb, Croatia

a (4, 2)' c

(2, 1)
(3, 1) (4,2)
k (4, 5)'
(3, 1)
f (5, 2) g (1, 3)

(1, 4) (3, 2) o
(2, 5)
Non-terminal n
nodes: [f k s r] (1, 1)

Representation: (2, 1)
s
s = [1 1 1 1] (4, 3)
1: included in solution (3, 3)
0: removed from graph q
(4, 1) (6, 1) (2, 1)
r
u (4, 2)
Zavod za telekomunikacije t 41
4
2
Fitness decoder function
University of Zagreb, Croatia

 Greedy delay constrained (3, 4) Δ=6

spanning tree algorithm for
a (4, 2)' c
given Steiner nodes
(2, 1)
 Start building tree at
(3, 1) (4,2)
source node
k (4, 5)'
 Select the node closest to (3, 1)
existing tree which does f (5, 2) g (1, 3)
not violate the delay
constraint (1, 4) (3, 2) o
(2, 5)
 If infeasible, fitness->∞ n
(1, 1)

(2, 1)
Non-terminal s
(4, 3)
nodes: [f k r s] (3, 3)
q
(4, 1) (6, 1) (2, 1)
Initial solution: r
u (4, 2)
s = [1 1 1 1] t
Fitness = 26

Zavod za telekomunikacije 42
4
3
Solving the problem using GRASP
University of Zagreb, Croatia

source (3, 4)

a (4, 2)' c Δ=6

(2, 1)
(3, 1) (4,2)
k (4, 5)'
(3, 1)
f (5, 2) g (1, 3)

(1, 4) (3, 2) o
(2, 5)
n
(1, 1)

(2, 1)
s
(4, 3)
(3, 3)
q
(4, 1) (6, 1) (2, 1)
r
u (4, 2)
t

Zavod za telekomunikacije 43
 GRASP Construction Phase
University of Zagreb, Croatia

source
(3, 4) Δ=6
a (4, 2)' c Unconnected
delay from
destination cost
(2, 1) source
nodes
(3, 1) (4,2)
k (4, 5)'
c 3 4
f (5, 2) g
(3, 1)
(1, 3)
g 4 2
o 3 4
(1, 4) (3, 2)
(2, 5)
o
n 4 5
n
(1, 1) t 6 6
(2, 1)
s
q 6 6 44
(4, 3)
q
(3, 3) u 13
10 76
(4, 1) (6, 1) (2, 1)
r
u (4, 2)
t 44
Zavod za telekomunikacije
 GRASP Construction Phase
University of Zagreb, Croatia

unconnected
destination cost
delay from RCL
source
nodes

c 3 4
cmin= 3, cmax= 13
g 4 2 α = 0.25 →
o 3 4 Threshold for RCL:
cmin+ α(cmax-cmin)=
n 4 5
3 + 2.5 = 5.5
t 6 6
q 6 6
u 13 6
Zavod za telekomunikacije 45
 GRASP Construction Phase
University of Zagreb, Croatia

(3, 4) update candidate list

a (4, 2)' c

(2, 1)
unconnected
delay from
destination cost
(3, 1) (4,2) source
nodes
k (4, 5)'
(3, 1) c 3 4
f (5, 2) g (1, 3)
3 2
g 4 2
(1, 4) (3, 2) o
(2, 5) n 4 5
n
3 6
(1, 1) t 6 6
(2, 1)
s
(4, 3) q 6 6
(3, 3)
q 12 6
(4, 1) (6, 1) (2, 1) u 13 6
r
u (4, 2)
t
Zavod za telekomunikacije 46
 GRASP Construction Phase
University of Zagreb, Croatia

unconnected
destination cost
delay from RCL
source
nodes

c 3 4 cmin= 3, cmax= 12

g 3 2 α = 0.25 →
Threshold for RCL:
t 3 6
cmin+ α(cmax-cmin)=
n 4 5 3 + 2.25 = 5.25
q 6 6 Procedure
continues until a
u 12 6 feasible solution
is built…

Zavod za telekomunikacije 47
4
8
GRASP Construction Phase
University of Zagreb, Croatia
(3, 4)
Note, there is a
a (4, 2)' c
cycle (this is not
(2, 1) an MST)
Cost=25
(3, 1) (4,2)
k (4, 5)'
(3, 1)
f (5, 2) g (1, 3)

(1, 4) (3, 2) o
(2, 5)
n
(1, 1)

(2, 1)
s
(4, 3)
(3, 3)
q
(4, 1) (6, 1) (2, 1)
r
u (4, 2)
t
Zavod za telekomunikacije 48
 Local Search Phase
University of Zagreb, Croatia
(3, 4)
Representation: set of Steiner (non-
a (4, 2)' c
terminal) nodes S represented by a
binary vector of size |V/D|: (2, 1)
1: included in solution (3, 1) (4,2)
0: excluded from solution k (4, 5)'
(3, 1)
f (5, 2) g (1, 3)
Non-terminal nodes: [f k r s]
Solution from GRASP construction (1, 4) (3, 2) o
phase: [1 1 0 1] (2, 5)
n
(1, 1)

Decoder function: find a DCST (Delay (2, 1)

s
(4, 3)
Constrained Spanning Tree) of sub- (3, 3)
q
graph G composed only of nodes S and (6, 1) (2, 1)
(4, 1)
D using a greedy approach (minimize r
u (4, 2)
cost while satisfying the delay t
constraint)
Fitness function: size of the found DCST
Zavod za telekomunikacije 49
5
0
Local Search Phase
University of Zagreb, Croatia

(3, 4)
Δ=6
Solution from construction (4, 2)' c
a
phase:
(2, 1)

Non-terminal nodes: [f k r s] (3, 1) (4,2)

k (4, 5)'
Solution Representation (3, 1)
s = [1 1 0 1] f (5, 2) g (1, 3)

Solution: (1, 4) (3, 2) o

the DCST shown (2, 5)
n
Fitness function: (1, 1)

f(s)=24 (2, 1)
s
(4, 3)
(3, 3)
q
(4, 1) (6, 1) (2, 1)
Note: cycle was u
r
(4, 2)
removed in t
decoding function

Zavod za telekomunikacije 50
5
1
Local Search Phase
University of Zagreb, Croatia

Current incumbent
N(s)
Selected solution
Current
nbr.
Solution: Current
Iteration Cost solution
s Nbr. incumbent Cost:
of s´
sol. Solution: c(sbest)
MST
sbest
Initialization: 1101 24
0101 21
1001 ∞
1 1101
1111 30
0101 0101 21
1100 ∞
1101 24
0001 ∞
2 0101
0111 27
1101 0101 21
0100 ∞

Max iterations without improvement = 1 (reached first local optimum)

Zavod za telekomunikacije 51
 The solution obtained at the end of the first
GRASP iteration
University of Zagreb, Croatia

(3, 4)
Δ=6
a (4, 2)' c

(2, 1)
(3, 1) (4,2)
Non-terminal nodes: [f k r s] k (4, 5)'
(3, 1)
Solution Representation f (5, 2) g (1, 3)
s = [0 1 0 1]
(1, 4) (3, 2) o
Solution: (2, 5)
the DCST shown n
(1, 1)
Fitness function: (2, 1)
s
f(s)=21 (4, 3)
(3, 3)
q
(4, 1) (6, 1) (2, 1)
r
u (4, 2)
t

Zavod za telekomunikacije 52