Metaheuristic Algorithms for Frequency Assignment Problems

D.K. Gupta Department of Mathematics Indian Institute of Technoiogy Kharagpur-721302,India E-mail: dkg@maths.iitkgp.ernet.in
Abstract
Frequency Assignment Problems (FAPs) are among the hardest combinatorial optimization problems of great practical importance in radio communication industries. They arise in practice when an area is to be served by a number of radio transmitters, or a network of radio links is io be established. Each transmitter or link is to be assigned a frequency so that the interference due to unwanted signals from transmitters on receivers has to be acceptable. The frequency assignment is also need to comply with certain regulations and physical characteristics of the transmitters. Moreover, the number of frequencies used in the assignment should be minimized. The aim of this paper is to experimentally investigate two metaheuristic algorithms namely Simulated Annealing ( SA ) and Tabu Search (TS) for solving fixed frequency assignment problems. Here, for the given fixed number of frequencies, we find the assignment to all the transmitters such that the fitness cost function E is minimized. E is formulated so that the number of violated constraints (ei,,), the sum of the amount by which each constraint is violated (erm ), the difference between the largest frequency (Garge ) and the smallest frequency (fsmail ) used, the number of distinct frequency used (eorder) the largest of the constraint violation, ( and ) can be minimized. Both the algorithms are tested on random FAPs with varying number of transmitters and available fiequencies. In almost all cases, SA performed better than TS measured in terms of E in getting optimized solutions

lhtroduction
The radio spectrum is an important natural resource. Many services, both civil and military, require access to the spectrum in order to function and the demand for frequencies is increasing year by year, Consequently, efficient management of the spectrum requires that fiequencies be assigned, for a particular service, in an optimum or near optimum manner. FAPs are NP Complete problems of great practical importance in radio communication industry. A radio communication network consists of a set of transmitters in a given region, where each transmitter has a fixed position in the region and an output signal which is to be assigned a single frequency fiom a set of uniformly spaced channels or discrete channels. The frequencies are assigned such that the observed interference strength at any position in the region is significantly less than the observed signal strength of the wanted signal at that same position, that is signal to noise ratio i s acceptable. Many other services such as cellular radio networks with homogeneous topologies and military radio link systems [4,5,6] non homogeneous topologies also require access with to this spectrum in order to function smoothly. Two types of problems generally occur are minimum span pnd fixed span problems. Minimum span frequency assignment involves assigning frequencies to a number of transmitters subject to a number of constraints such that no interference is suffered and the span of frequencies used, i.e., the difference between the largest and smallest frequency used, is minimized. In contrast, fixed

spectrum frequency assignment involves assigning frequencies from a fixed set of frequencies. Many researchers [ 1,3] have studied a number of approaches for efficient management of frequency spectrum so that frequencies can be assigned for a particuIar service in an optimal or near optimal manner. Exact methods based on graph theory gives optimum solution but are computationally unrealistic. The objective of this paper is to experimentally investigate two metaheuristic algorithms namely Simulated Annealing ( SA ) and Tabu Search ( TS ) for solving fixed frequency assignment problems. Here, for the given fixed number of frequencies, we find the assignment to all the transmitters such that the fitness cost function E is minimized. E is formulated so that the number of violated constraints (eno), the sum of the amount by which each constraint is violated (esum >, the difference between the largest frequency (fiarg, ) and the smallest frequency (fsmall used, the number of distinct frequency ) used (eordcr) and the largest of the constraint violation, ( I,, ) can be minimized. E is given by E = Pi eno + PZerum + ~3 (f;BTpC - L i i ) + ~4 eordFr + p5 f;arge + F16 IMO where, pi are weights, which can include 0, reflecting the relative importance of the various factors. Both the algorithms are tested on random FAPs with varying number of transmitters and available frequencies. For both SA and TS, the starting solution which is a no violation assignment is generated either randomly or by sequential algorithms [Z]. Two generators called single move (assignment differing in one component) and double move ((assignment differing in two components ) can be used to produce new assignment from the current assignment. For SA, the starting temperature' is taken as 1 and two types of annealing schedules are used. The first one is the geometric cooling-lgiven by tk+l = CI tk with a E [O,I]. The scheme decreases the temperature by the specification of the parameter a. The number of iterations at each step is fixed at N. The second one cooling-2 scheme uses the number of iterations at iach temperature given by Nk+l= Nk/ a , where No = N is the number of transmitters and the temperature reduction is again specified by the parameter CL . As the temperature decreases, the number of trial assignments tested at each temperature increases. SA is terminated when the final temperature falls to user specified minimum temperature. In TS, the search space of all feasible solutions is explored by a

456

sequence ofmoves. A move from one solution to other solution is best available. However, to escape from IocalIy optimal but not globally optimal solutions and to prevent cycling, some moves at one particular iteration, are classified as forbidden or tabu: Tabu moves are based on the short term and long term history of the sequence of moves. Both the algorithms are tested on random FAPs with varying number o f transmitters and available frequencies. In almost all cases, SA performed better than TS measured in terms of E in getting optimized solutions This paper is organized as folfows. In section 2, we describe in brief the definition, concepts and notations used in this paper to make it self explanatory. Section 3, discusses in brief SA in solving FAP and experimental results obtained for it. Section 4,gives the TS in brief and its experimental results. A comparative evaluation of SA and TS forms the section 5. Finally conclusion are included in section 6.
2. Definitions and Notations In order to model interference, constraints are imposed on the assignment. When a transmitter can cause interference with a receiver, a constraint is implied on the frequencies assigned to this transmitter and the transmitter serving the receiver Interference can occur when the assigned frequencies are the same or close together. This can happen when the transmitters and receivers are at the same location or within a few tens of meters of each other (co-site interference), or, when they are at a distance of several kilometers or more (far-site interference). Constraints are of the following types. CO-channel constraints: A pair of transmitters located at different sites must not be assigned the same frequency, unless they are sufficiently geographically separated. If f, and fJ are the frequencies assigned to transmitters i and j then this gives rise to constraints of the form.

There are (m)N possible frequency assignments, and because it may be difficult, or even impossible, to determine one which imposes no interference we determine an assignment where the total interference is minimal. I f f = (fl, f 2 , .. .. fN) denotes a frequency assignment and E(f) denotes the total interference, the problem is to determinean f E DN for which E is minimal or near minimal.

3. Simulated Annealing For FAP

Simulated annealing (SA) is a stochastic computational technique derived from statistical mechanics For finding near global-minimumcost solutions to large optimization problems. The method itself has a direct analogy with thermodynamics, specifically with the way that liquids freeze and crystallize, or metals cool and anneal. At high temperatures, the molecules of a liquid move freely with respect to one another. If the liquid is cooled slowly, thermal mobility is restricted. The atoms are often able to line themselves up and form a pure crystal that is completely regular. This crystal is the state of minimum energy for the system, which would correspond to the optimal solution in a mathematical optimization problem, However, if a liquid metal is cooled quickly i.e. quenched, it does not reach a minimum energy state but a somewhat higher energy state corresponding, in the mathematical sense, to a suboptimal solution. SA will find a solution X = (xi, x ? . .... XN ) and will require agenerator of random changes in a configuration, a cost function E(X) whose minimization is the goal , a control parameter t (analogue of temperature) and an annealing schedule which indicates how t is lowered from high values to low values, e.g., after how many random changes in configuration is t reduced and by how much? Given a succession of moves SA was assumed to change its configuration from energy p = Eold to q = E,, with probability prob = - ( P - 4 ) I I ) where, B is known as Boltzman constant. If E,,, < Eold then the new configuration has a lower energy state than the old one and the system always accepts this move. If E > Eoldthen the , new configuration may still be accepted with probability prob (0 prob 1) and thus help the system jump out of a local minimum. Let NUMI,,, is the number of random changes in configuration at each temperature t and is chosen so that the configuration has reached a minimum energy state for the current temperature.

L +f,
Adjacent channel constraints: When a transmitter and a receiver are tuned to similar frequencies, there is still the potential for interference. Therefore a number of constraints arise of the following form

M-f,I>m
for some value of m, where m is the required number of channels

separation.
Co-site frequency separation: Any pair of frequencies at a site must be

separated by a certain fixed amount of channel. If a frequency is to be used by a high power transmitter then its frequency separation should be larger, say 500 kHz or 10 channels. The constraint can therefore be of the form:

IL-f,

12m

where m refers to the number of channels separation required between transmitters i and j. For a given problem containing N transmitters, numbered from 1 to N, these interference constraints are represented by a N x N constraint matrix correspanding to a constraint graph with N vertices. The Fixed spech-uT model used in the paper is given as follows. A fixed set of m frequencies is denoted by: D I di, d2, ..-,4 ) It is required to determine N values
(fly f2,

.... f N )
b C U

N such that, for all i; j with 1 I< j I i

I c-f,

457

SA Algorithm

Initialise t Generate random configuration Xold WHILE t > tmi, DO FOR i = I to NUMloop DO generate new configuration, X,,, , from calculate new energy,

X, ,

Eve,+

calculate AE = E,, -Eo,d

IF AE < 0 or random
X",d

-=prob = C"''

THEN

x, ,

Eold

5 E",,

END IF END FOR reduce t (e.g. t = 0:9t)

END WHILE
The cost function to be minimized is given by E = PI e G o + ,u2esum F I (fiarge - L a i i 1+ ~4 eordcr PSfiargc + P Llvio + ~ + ~ Two different generators, single move and double move [ 21 can be used to produce a new assignment from the current assignment. The starting temperature is determined by first setting to = 1 and performing 100 iterations of the FOR loop from the main loop of the above algorithm. If the acceptance ratio, x, defined as the number of accepted trial assignments divided by 100 (NUMloop), is less than 0.9, double the current value of to. Continue this procedure until the observed acceptance ratio exceeds 0.9 (with reinitialised to zero prior to starting the FOR loop). The algorithm terminates when the temperature, fk , falls below

are classified as forbidden or tabu (or taboo). Tabu moves are based on the Short -term and long-term history of the sequence of moves. A simple implementation, for example, might classify a move as tabu if the reverse move has been made recently or frequently. Sometimes, when it is deemed favorable, a tabu move can be overridden. Such aspiration criteria might include the case which, by forgetting that a move is tabu, leads to a solution which is the best obtained so far. Suppose h is the real valued objective function on a search space S such that h(s) has minimal or near minimal value. Ts ,commences with a possibly random solution so E S and determines a sequence of solutions so. sl,.,., s, E S. At each iteration, s,.~( 0 5 j < n) is selected from the neighborhood N,,, (s,) . The process of selection is first to determine the tabu set T,,, (s,) r N,,, (sj) of neighbors of sj and the aspirant set A,,, (s,) E T,,, (sj) of tabu neighbors. The sj+]is the neighbor of sj which is either an aspirant or not tabu and for which h(sj+,) is minimal. That is, h(sj+l)5 h(4) for all si E w e t (SI - T,,, (SI) A, (SI. s,

fmin (user specified) or the number of frozen temperatures exceeds a user specified value, usually IO. A frozen temperature occurs when no new assignments are accepted for a given temperature t , (i.e. after NUMloop

iterations of the FOR loop). Three annealing schedules are included in the Simulated Annealing. In all three cases the parameter NUMloop is set to N, the number of transmitters. Two annealing schedules used are Cooling-I, where the cooling scheme tk is reduced according to the formula f k + , =atk; with a f [O; 1 3 . In the cooling-2 scheme, the temperature reduction is again specified by the parameter a, however the number of iterations at each temperature is specified by
Nk+l =

[Nd a 1

(up to a maximum of 2000N), where No = N,the number of transmitters. As the temperature decreases the number of trial assignments tested at each temperature increases. This scheme was found to be particularly effective. 4.Tabu Search for FAP. Tabu search (TS) was first suggested by Glover [SI and since then has become increasingly used. It has been successfully applied to obtain optimal or sub-optimal, solutions to such problems as scheduling, timetabling, traveling salesman and layout optimization problems. The basic idea of the method, described by Glover, Taillard and de Werra [ 9 ] , is to explore the search space of all feasible solutions by a sequence of moves. A move from one solution to another is the best available. However, to escape from locally optimal but not globally optimal solutions and to prevent cycling, some moves, at one particular iteration,

Tabu Search (TS) Algorithm Generate initial solution s WHILE not finished Identify N,,, (s) c S (neighborhood set) Identify T,,, (s) E N,,, (s) {tabu set) Identify A,,, (s) c T,,, (s) (aspirant set) Choose si E (NSet( s ) - T,,, (s)) U A,,, (s) , for which h(s') is minimal. s= s/ END WHILE Let the frequency assignment f = ( f,, fi, ..., fN) is represented using an arrayof indexes [x,, x2,... xN 1, where fj =dxj for 1 I j 5 N. The full neighborhood includes all the neighbors of an assignment f produced by using the single move generator as described far SA. Thus, for TS this involves calculating N ( Ffixcd ) neighbors at -1 each iteration. A move to a neighbor (i, x/) corresponding to changing the assignment of the transmitter I to x/ is said to be tabu if it does not satisfy the short-term or long-term memory condition. These conditions are determined by two numbers: a positive integer r and a real number p, where 0 < p < I . The short -term memory condition specifies that any transmitter I can not be assigned the same frequency over any of the previous r moves. The long-term memory condition specifies that the proportions of the number of times transmitter 1 had been changed over all iterations does not exceed 8.

458

Thus, if at iterationj the frequency dxj' asigned to tmsmitter I has changed than a move at iteration k+l is tabu if either short-term memory dxiik+' = dxj' where k-r t 5 k At each iteration the method selects from the nontabu neighbors that neighbor fik+')or f) for which @ > E (ek) ). This aIlows escape from local minima. Ts allows a tabu move to be selected when certain aspiration criteria are satisfied. In our implementation a tabu move is selected if the neighbor fik+''satisfies E(Pk*')) < E (I!) for all neighbors f' of 6k) and < E ( P ) for all I 5 j 5 k, To summarize, if ftabu denotcs the best tabu neighbor o f tk) and f,,, ,the best nontabu neighbor, then the rule for determining f @ + l ) is if E ( f ~ , ) E (Imn) E(fmbu < E ($) ) for 1 5 j 5 k then pk" = cabu and ) else P = f,,, I Note that there are some implicit relations between the parameters r and p that have to be satisfied. If r > N or p <I/", then after N nontabu moves, every move is tabu. Also,~after iterations of nontabu moves, it follows k from the recency condition that a transmitter can have changed at most rkirl times. Hence the long-term memory value, to have any effect, should satisfy kp < rkirl We therefore obtain a relation between r and p which is 1/N $<l/r, In practice, we choose p = h/r + (I-h)/N for some value of h with 0 1.

Schedule Two Type

15 1'2
17
1 2

~57228419
I

156

559.488 7
122439
1 0

63

(59851717 + 821.784 7
892758
7

159
86 95
49
8 9

1 9

1 2 1 4
1 4

83

15
1 7 19

449.766 9
594444
9

50
56

467%
815395

14

123437

10
I

83

111355 8

Dl

5. Results and Discussions. The follwing results are obtained from SA and TS for solving a number of FAPs for different number of transmitter (N)and available frequencies (MaxF).

t 15

. .

12

I 427.51

I9

I 49 I 614.527 I 6

16 5

One can easily see that SA performs better than TS.

References. [ I ] S.U.Thie1, S.Hurley and DSmith; Frequency Assignment Algorithms Radiocommunications Agency Agreement Ref. RCCm 070 (1 997) 1-80. [Z] W.K.hale; Frequency assignment:theory and applications, Proc.lEEE,68,(19SO)1497-15 14. [3] S.hurley, D.H.Smith and S.U. Thiel; FASoft: A system for discrete channel frequency assignment, Radio Science, Vol-32 ( 1 997)192 1-1 939. [4] Bouju, A., J.F.Boyce, C.H.D. Dimitropoulos, G. vom Scheidt and J.G. Taylor; Tabu search for the radio Iinks frequency assignment problem, in conference on Applied Decision Technologies: Modem Heuristic methods, (1995) 233-250. [5] S.U.Herley, and D.H.Smith; Fixed Spectrum Frequency assignment using natural algorithms, in Genetic Algorithims in Engineering Systems, (1995) 373-378 [6] L.G. Anderson; A simulationstudy of some dynamic channel aassignment algorithms in a high capacity mobile communications system, IEEE trans. Commun., COM-21 (1973)12941301 [7] S. Kirkpatrick, C.D.Gelatt, M.D.Vecchi; Optimization by simulated Annealing, Science, 220 (1983) 671-610. [XI F.Glover; Tabu Search Part4 & 11, ORSA Journal on Computing,I & 2, (1989 ) and ( I 990). [9] F. Glover, E. Taillard and D.de Werra; A user's guide to tabu search, Ann. Oper. Res, 4 1 (1 993) pp3-28

459