Sei sulla pagina 1di 55

Optimization Techniques for Telecommunications

Universit degli Studi di Trieste

Lorenzo Castelli

ICTP ITU School on New Perspectives in Wireless Networking Trieste, 27 February 2008

Outline
Introduction to Optimization A simple case study Basics in Linear Optimization Simplex Algorithm Integer Programming References and Resources A research topic
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Travelling Salesman Problem (TSP)


For a set of nodes and a set of arcs connecting such nodes, identify the minimum cost cycle that visits each node only once
6 3 2 4 7 6 2 5 4 5 6 3
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

9 3 7 4 6 8

2 7

3 9 4

2 6

3 5

Travelling Salesman Problem (TSP)


For a set of nodes and a set of arcs connecting such nodes, identify the minimum cost cycle that visits each node only once
6 3 2 4 7 6 2 5 4 5 6 3
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

9 3 7 4 6 8

2 7

3 9 4

L = 71

2 6

3 5

Travelling Salesman Problem (TSP)


For a set of nodes and a set of arcs connecting such nodes, identify the minimum cost cycle that visits each node only once
6 3 2 4 7 6 2 5 4 5 6 3
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

9 3 7 4 6 8

2 7

3 9 4

L = 62

2 6

3 5

An Optimization Problem
Let S " ! n be the Feasible Set or the Set of feasible solutions Let f : S a ! be a real function to be optimized, i.e., maximized or minimized. It is called objective function The solution of an optimization problem is to find the set of optimal solutions, i.e.,
OptSolM = s * ! S : f ( s * ) # f ( s ) "s ! S

OptSolm

{ } = { ! S : f ( s ) # f ( s ) "s ! S } s
* *
L. Castelli University of Trieste

New Perspectives in Wireless Networking ICTP, 27 Feb 08

Numerical Example
max f(x,y) = x + y S = {x 0; y 0; x 4; y 2; 3x + 5y 15} y
3

x* = 4.0 y* = 0.6

S
1 2 3 4

s* x
5

f(x*,y*) = 4.6

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Optimization characteristics The objective function can be linear or non-linear (e.g., quadratic) continous or not (e.g., fixed cost) differentiable or not (e.g., piecewise linear) The feasible set can be convex or not

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Linear Programming (LP)


f is a linear function S " ! n can be represented by a set of m linear inequalities Any LP problem can be formulated as max cTs As b
where dim A = m x n; dim b = m x 1; dim c = n x 1; all values are deterministic
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

s0

Integer Programming (IP) We define IP when all elements in S assume only integer (or binary) values If some elements in S assume integer (or binary) values, whereas other elements assume continous values, we refer to as Mixed Integer LP (MILP)
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Optimization process Analyse the system to optimize Formulate the model Solve the model by appropriate algorithms

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Optimization process Step 1

Analysis of the system to be optimized

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Location of GSM transmitters


A mobile phone operator decides to equip a currently uncovered geographical zone The management allocates a budget of 10 million to equip this region A study shows that only 7 locations are possible for the construction of the transmitters It is also known that every transmitter only covers a certain number of communities
Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Geographic specifications
A schematic map of the region with the division into communities and the possible locations for transmitters Every potential site is indicated by a black dot with a number, every community is represented by a polygon. The number in the center of a polygon is the number of the community

Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

GSM transmitter specifications


Certain geographical and topological constraints add to the construction cost and reduce the reach of the GSM transmitters. We list the communities covered and the cost for every site.
Site 1 2 3 4 5 6 7

Cost (in Million )

1.8

1.3

4.0

3.5

3.8

2.6 7,10,11,12,1 5

2.1 12,13,14,1 5

Communities covered

1,2,4

2,3,5

4,7,8,10

5,6,8,9

8,9,12

Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Data on population

For every community the number of inhabitants is known

Community

1 2

4 5 6 7

10 11 12 13 14 15 6 14 9 3 6

Population (000) 2 4 13 6 9 4 8 12 10 11

Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Management dilemma

Where should the transmitters be built to cover the largest population with the given budget limit of 10M?

Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Optimization process Step 2

Model Formulation

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Formulation steps
Identify the decisions to be taken which are the variables of what type (i.e., continous, integer, binary) Define the objective function Determine the constraints Identify the available data, i.e., matrix A and vectors b and c
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Identify the variables Two types of binary variables are required variables xc that are 1 if and only if a community c is covered by a transmitter variables yp that are 1 if and only if a transmitter is built at site p

Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

The available data


SETS COMMS the set of communities PLACES the potential sites for constructing transmitters Vector c POPc the number of inhabitants of the community c Vector b BUDGET the budget allocated by the management Matrix A COSTp for the cost of building a transmitter a site p COVERpc a binary constant that indicates whether a transmitter placed at p covers community c (COVERpc = 1) or not (COVERpc = 0)
Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

The objective function


To maximize the population covered by the new GSM transmitters

max

c"COMMS

! POP # x
c

Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Budget constraint
It is necessary to remain within the budgetary limits, i.e., the total construction cost for transmitters has to be not greater than the given maximum budget

p"PLACES

! COST

$ y p # BUDGET

Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Coverage constraints (I)


We need to translate the equivalence community c receives a GSM signal iff at least at one site covering this community a transmitter is built, or xc = 1 iff there is at least one p with COVERpc yp = 1 BUT this type of equivalence cannot be translated directly into a linear form Since a community may be covered by more than one transmitter, the constraints

p!PLACES

% COVER

pc

$ y p # xc

"c ! COMMS

express one direction of the equivalence by specifying that the sum COVERpc yp is greater than or equal to xc
Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Coverage constraints (II) The other direction of the equivalence is ensured through the maximization of the population covered The optimization algorithm will not leave any xc at 0 if any transmitter that covers this community is built

Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Binary constraints
All the variables are binary

xc ! { ,1} "c ! COMMS 0


y p ! { ,1} "p ! PLACES 0

Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Complete model formulation


max

c"COMMS

! POP # x
c
p

p"PLACES

! COST
pc

$ y p # BUDGET
"c ! COMMS

p!PLACES

% COVER

$ y p # xc

xc ! { ,1} "c ! COMMS 0


y p ! { ,1} "p ! PLACES 0
Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Problem size 15 communities, 7 sites 15 + 7 = 22 binary variables 222 = 4 194 304 15 + 1 = 16 constraints dim A = 16 x 22
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Exercise Formulate an IP model to determine the GSM transmitters location of minimum cost such that ALL the communities are covered No budget constraint

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Exercise Solution
min

p!PLACES

COSTp " y p

p!PLACES

COVER pc " y p # 1 $c ! COMMS

y p ! { ,1} "p ! PLACES 0

Set Covering Problem

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Optimization process Step 3

Development of solving algoritms

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Computational Complexity If all the variables are real, it is easy to solve the problem, i.e., it exists an algorithm running in polinomial time When integer or binary variables, we may face difficult problems, i.e., NPhard

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Linear Programming (LP)


f is a linear function S " ! n can be represented by a set of m linear inequalities Any LP problem can be formulated as max cTs As b
where dim A = m x n; dim b = m x 1; dim c = n x 1; all values are deterministic
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

s0

Solving LP problems A LP problem is easy to solve Simplex algorithm Interior point method Most used is the Simplex algorithm Theoretically it is not a fast algorithm, but it is solved quickly in practice
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

The simplex algorithm The feasible set is a convex set The optimal solution, if it exists, lies on a vertex of the feasible set At each step of the algorithm the objective function does not decrease (if maximization) or does not increase (if minimization)
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Simplex Algorithm - Numerical Example


max f(x,y) = x + y S = {x 0; y 0; x 4; y 2; 3x + 5y 15} x* = 4.0 y* = 0.6 f(x*,y*) = 4.6

y
3

A
2

Two possible paths 1. O D C 2. O A B C


C

s*
D

We move from one vertex to an adjacent one, without worsening the objective function
5 L. Castelli University of Trieste

New Perspectives in Wireless Networking ICTP, 27 Feb 08

Simplex Algorithm - Stopping criteria

The simplex algorithm detects If the feasible set is empty or not If not empty If the set of optimal solution is empty or not If not empty If there is a unique or infinite solutions
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Simplex Algorithm Computing time

Even large istances can be solved in a few seconds with freely available or commercial software E.g., Flux Balance Analysis to model fluxes in a metabolic network of a cell: 800 rows and 1200 columns

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Integer Programming The feasibility set is no longer convex What is a vertex of the feasible set?
y
max f(x,y) = x + y
3

A
2 1

s*
C D

S = {x 0; y 0; x 4; y 2; 3x + 5y 15} x, y integer s* = (4,0) or s* = (3,1)

f(s*) = 4
5 L. Castelli University of Trieste

New Perspectives in Wireless Networking ICTP, 27 Feb 08

IP The Feasible Set Differently from LP, the IP feasible set is composed of a finite number of points It may turn out to be unworkable to enumerate them all Alternative algorithms are required

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

IP Solving Algorithms
Exact Algorithms (e.g., Branch&Bound, Gomory cuts) reach soon or later the optimal solution may run in exponential time Approximate Algorithms (e.g., Christofides for TSP) run in polynomial time do not necessarily reach the optimum the gap between the optimum and the found solution is bounded Heuristics run in polynomial time do not necessarily reach the optimum the gap between the optimum and the found solution is unknown Metaheuristics Genetic Algorithms, Tabu Search, Simulated Annealing

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Global and local optima


Local Optima

GAP

Global Optimum
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Implementing the algorithms Simplex and exact MIP algorithms are available through ad-hoc optimization software XPress-MP, ILOG CPLEX, GLPK, Approximate and heuristic algorithms can be written in any language, as C, C++, FORTRAN, Libraries, interfaces are available
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Implementation XPress-MP (I)


model "G-6 Transmitter placement" uses "mmxprs" declarations COMMS = 1..15 ! Set of communities PLACES = 1..7 ! Set of possible transm. locations COST: array(PLACES) of real ! Cost of constructing transmitters COVER: array(PLACES,COMMS) of integer ! Coverage by transmitter locations POP: array(COMMS) of integer ! Number of inhabitants (in 1000) BUDGET: integer ! Budget limit build: array(PLACES) of mpvar ! 1 if transmitter built, 0 otherwise covered: array(COMMS) of mpvar ! 1 if community covered, 0 otherwise end-declarations initializations from g6transmit.dat COST COVER POP BUDGET end-initializations
Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Implementation XPress-MP (II)

! Objective: total population covered Coverage:= sum(c in COMMS) POP(c)*covered(c) ! Towns covered forall(c in COMMS) sum(p in PLACES) COVER(p,c)*build(p) >= covered(c) ! Budget limit sum(p in PLACES) COST(p)*build(p) <= BUDGET forall(p in PLACES) build(p) is_binary forall(c in COMMS) covered(c) is_binary ! Solve the problem maximize(Coverage) end-model
Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Location of GSM transmitters - Solution Total coverage: 109 Total cost: 9.5 Build transmitters: 2 4 6 7 Communities covered: 2 3 5 6 7 8 9 10 11 12 13 14 15 Communities 1 and 4 are NOT covered
Source: Applications of Optimization, Chap. 7 www.dashoptimization.com New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Set Covering Problem - Solution All Communities covered Total cost: 11.3 Build transmitters: 1 2 4 6 7

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

Resources www.informs.org www.scienceofbetter.org Handbook of Optimization in Telecommunications www.scopus.com scholar.google.com


New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Last Published Papers


SEARCH in www.scopus.com with TELECOMMUNICATIONS & OPTIMIZATION
1. 2. 3. 4. 5. 6. 7. 8. 9. Rerouting tunnels for MPLS network resource optimization Klopfenstein, O. 2008 European Journal of Operational Research 188 (1), pp. 293-312 A multi-criteria approach to fair and efficient bandwidth allocation Ogryczak, W., Wierzbicki, A., Milewski, M. 2008 Omega 36 (3), pp. 451-463 Optimizing revenue in CDMA networks under demand uncertainty Olinick, E.V., Rosenberger, J.M. 2008 European Journal of Operational Research 186 (2), pp. 812-825 Benders decomposition for the uncapacitated multiple allocation hub location problem de Camargo, R.S., Miranda Jr., G., Luna, H.P. 2008 Computers and Operations Research 35 (4), pp. 1047-1064 Multi-period capacitated location with modular equipments Gourdin, E, Klopfenstein, O. 2008 Computers and Operations Research 35 (3), pp. 661-682 Developing staff schedules for a bilingual telecommunication call center with flexible workers Ertogral, K., Bamuqabel, B. 2008 Computers and Industrial Engineering 54 (1), pp. 118-127 Optimum VoIP packet routing using a communications processor Zupanc ic , S., Polans ek, G., Z emva, A. 2008 AEU - International Journal of Electronics and Communications 62 (2), pp. 104-113 Max-min fairness in multi-commodity flows Nace, D., Nhat Doan, L., Klopfenstein, O., Bashllari, A. 2008 Computers and Operations Research 35 (2), pp. 557-573 Volterra-series-based distortion analysis for optimization of out-of-band terminations in GaN HEMT devices Srinidhi, E.R., Ma, R., Kompa, G. 2008 IEEE Electron Device Letters 29 (1), pp. 24-27

New Perspectives in Wireless Networking ICTP, 27 Feb 08

L. Castelli University of Trieste

A joint research ARPL Lab - UniTS Minimum cost placement of WI-FI access points considering Directional antennas Client coverage The presence of obstacles Overlapping channels We have to decide with type of antenna and in which location
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Interference Graph
There is an arc if two antennas interfere when using the same channel Only 3 channels available (1,6,11) A feasible solution is a 3-colour graph
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Interference Graph
There is an arc if two antennas interfere when using the same channel Only 3 channels available (1,6,11) A feasible solution is a 3-colour graph
L. Castelli University of Trieste

Feasible solution
New Perspectives in Wireless Networking ICTP, 27 Feb 08

Interference Graph
There is an arc if two antennas interfere when using the same channel Only 3 channels available (1,6,11) A feasible solution is a 3-colour graph
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Interference Graph
By taking away any point in the dotted square we may have a feasible solution Feasibitly requires that the dimension of the max clique is 3 (difficult problem, NP-complete)
L. Castelli University of Trieste

Unfeasible solution
New Perspectives in Wireless Networking ICTP, 27 Feb 08

Contact Details Lorenzo Castelli DEEI University of Trieste Via A. Valerio 10 34127 Trieste, Italy Email: castelli[at]units.it Url: www.units.it/castelli
New Perspectives in Wireless Networking ICTP, 27 Feb 08 L. Castelli University of Trieste

Potrebbero piacerti anche