238

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 15, NO. 2, FEBRUARY 1997

Static and Dynamic Channel Assignment Using Neural Networks

Kate Smith, Member, IEEE, and Marimuthu Palaniswami, Senior Member, IEEE
AbstractIn this paper, we examine the problem of assigning calls in a cellular mobile network to channels in the frequency domain. Such assignments must be made so that interference between calls is minimized, while demands for channels are satised. A new nonlinear integer programming representation of the static channel assignment (SCA) problem is formulated. We then propose two different neural networks for solving this problem. The rst is an improved Hopeld neural network which resolves the issues of infeasibility and poor solution quality which have plagued the reputation of the Hopeld network. The second approach is a new self-organizing neural network which is able to solve the SCA problem and many other practical optimization problems due to its generalizing ability. A variety of test problems are used to compare the performances of the neural techniques against more traditional heuristic approaches. Finally, extensions to the dynamic channel assignment problem are considered. Index Terms Channel assignment, combinatorial optimization, Hopeld neural network, self-organizing neural network.

I. INTRODUCTION

VER RECENT years, we have seen a steady increase in the popularity of cellular mobile communication systems. The fact that the electromagnetic spectrum available for this purpose is a limited resource places severe limitations on the size and performance of such systems. Careful design of a network is necessary to ensure efcient use of the available spectrum. A competitive communications industry also forces the network design to accommodate certain performance criteria. Operators of such networks need to assign a caller from a certain region to a channel within the frequency spectrum in such a way that the probability of the signal-tointerference ratio (SIR) being below some predened limit is low. Potential interferences to this call come from three possible sources: another caller within some range using the same channel (a co-channel interference); another caller in an adjacent region using an adjacent channel in the frequency domain (an adjacent channel interference); and another caller within the same region using another channel within some range (a cosite interference). The combinations of regions and channels which cause interferences are determined by the radio frequency (RF) propagation (obtained from the regional
Manuscript received February 1996; revised July 1, 1996. This work was carried out while K. Smith was a Ph.D. student at the University of Melbourne, Parkville, Victoria, Australia; she was supported by an Australian Postgraduate Award (Priority) and a CSIRO Supplementary Scholarship. K. Smith is with the Department of Business Systems, Monash University, Clayton, Victoria, 3168, Australia. M. Palaniswami is with the Department of Electrical and Electronic Engineering, The University of Melbourne, Parkville, Victoria, 3052, Australia. Publisher Item Identier S 0733-8716(97)00474-5.

topography and morphostructure) and the spatial density of the expected trafc. Specialized programs have been created for the purpose of calculating these relationships [20]. The expected trafc in a particular region can also be used to predict the demand for channels. The channel assignment problem (CAP) is then to assign the required number of channels to each region in such a way that interference is precluded and the frequency spectrum is used efciently. Clearly, there are two steps to achieving this assignment successfully. The rst is the design process, where future trafc is forecast and the assignments are made to satisfy the requirements for the immediate future. This is a static channel assignment (SCA) since the assignment of channels to regions is made only once and remains xed. As the trafc increases, however, it will become increasingly difcult to maintain performance standards with this network. A more dynamic assignment strategy will need to be employed which permits rearrangement of assignments in an on-going adaptive environment. The task of any dynamic channel assignment (DCA) procedure is dependent, to a large extent, on the success or otherwise of the initial SCA, and it is for this reason that a great deal of research has focused on this static design problem. It has been shown that the SCA problem is a generalized graph coloring problem [42] and is therefore NP-hard. Many heuristic techniques have been devised for solving the SCA problem [3], [9], [38]. Research has also been carried out on the theoretical components of the problem, including obtaining lower bounds for the number of channels necessary to obtain an interference-free assignment [18], [21]. Hopeld neural networks have been used to solve the SCA problem [17], [32], [33], although the types of energy function chosen are of the HopeldTank kind [29] which involve many terms and consequently result in infeasibility and poor solution quality. This is the conclusion reached by Kunz [33]. In this paper, we reformulate the SCA problem as a generalized quadratic assignment problem, treating the noninterference constraints as soft constraints in the objective function and the demand satisfaction constraints as hard constraints. The advantage of this approach is that a solution which minimizes the severity of any interferences will always be found, which is particularly useful if the demands and constraints are such that no interference-free solution exists. This reformulation is described in Section II. In Section III, we briey survey the existing techniques for solving the SCA problem, and describe simulated annealing and steepest descent heuristics which minimize our reformulated problem.

07338716/97$10.00 1997 IEEE

SMITH AND PALANISWAMI: STATIC AND DYNAMIC CHANNEL ASSIGNMENT USING NEURAL NETWORKS

239

An improved Hopeld network approach is then proposed in Section IV-A. The Hopeld energy function is condensed so that the constraints of the problem are contained in a single term, while the internal dynamics of the network are modied to permit hill-climbing and escape from local minima. In Section IV-B, a new self-organizing neural network is described which is a generalization of Kohonens self-organizing feature map [31], with signicant modications to enable optimization. Selection of the parameters involved in these novel neural network approaches to solving the SCA problem are also discussed. Two main sets of test data are used to compare the performance of these methods in Section V. The rst set of data consists of an articial network of 21 hexagonal cells, with variations in demands, and the noninterference constraints. The number of channels with which to nd the best assignment is xed according to the lower bound rules of Gamst [21]. The second data set is based upon a 24 21 km area around Helsinki, Finland, which is a common test set in the literature [32]. Results comparing the performances of the simulated annealing and steepest descent heuristics, a commercial optimization package, as well as the improved Hopeld network and new self-organizing neural network are presented and discussed. We then consider extensions to the dynamic rearrangement of channels in Section VI. II. STATIC CHANNEL ASSIGNMENT A. Problem Description cells (or regions) in the network and Suppose there are the number of channels available is given by . The channel requirements for cell are given by . Assuming that the RF propagation and the spatial density of the expected trafc have already been calculated, the noninterference constraints can be determined (perhaps using a program like GRAND [20]). The constraints specify the minimum distance in the frequency domain by which two channels must be separated in order that an acceptably low S/I ratio can be guaranteed within the regions to which these channels have been assigned. These minimum distances are stored in a symmetric compatability matrix which is of dimension . As an example, suppose the number of cells in the network is and the demand for channels in each of these cells is given by . Consider the compatability matrix suggested by Sivarajan et al. [38]

Fig. 1. An interference-free assignment for a four-cell and 11-channel network.

Fig. 2. A near interference-free assignment for a four-cell and 10 channel network.

The channel assignment problem, as demonstrated by this example, is then as follows (CAP1) minimize subject to span of channels demand and noninterference constraints

given and . It has been shown [38] that the minimum number of channels needed for an interference-free assignment using this example is 11. Thus, is the lower bound, and we will be unable to nd any interference-free assignments if . Suppose, however, that there are only ten channels available. There is little point trying to nd a solution which minimizes the span of the channels if the number of available channels is less than the lower bound. An alternative and more useful version of the channel assignment problem is then (CAP2) minimize severity of interferences subject to demand constraints

(1)

indicate that any two channels The diagonal terms assigned to cell must be at least ve frequencies apart in order that no co-site interference exists. Channels assigned to cells 1 and 2 must be at least frequencies apart. Off-diagonal terms of and correspond to cochannel and adjacent channel constraints, respectively. Thus, the solution represented in Fig. 1 is one of several interferencefree assignments which can be generated for .

given and . A solution to CAP2 for is shown in Fig. 2, where it can be seen that the chosen assignment creates co-site interference within cell 4, but the severity of this interference has been minimized and it is relatively weak (the channels are four apart, rather than the minimum of ve required for no interference). Clearly, it is immensely more useful in practical situations (where an unlimited number of channels is not available) to attempt to obtain the best assignments possible given the number of channels available. If the number of available channels is greater than the lower bound (the minimum number of channels required for an interference-free assignment), then CAP1 should be used to try to make efcient use of

240

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 15, NO. 2, FEBRUARY 1997

the available spectrum. Typically, however, the number of available channels is substantially less than the lower bound, in which case CAP2 should be solved. In the following sections, we formulate CAP2 as a 01 generalized quadratic assignment problem, which can be solved using both neural and heuristic techniques. B. Mathematical Formulation Let us assume there are cells and channels available for the network. We dene a set of binary variables if cell is assigned to channel otherwise and . Suppose that ; that is, calls in cells and have been assigned channels and , respectively. One way to measure the degree of interference caused by such assignments is to weight each assignment by an element in a cost tensor , where is the distance (in the channel domain) between channels and . If , then the interference cost should be at its maximum, with the cost decreasing until the two channels are far enough apart that no interference exists. The SCA problem can thus be formulated to minimize the total cost of all assignments in the network for minimize (2)
Fig. 3. Proximity factor tensor for example compatability matrix.

the strength of the resulting signal to still be as strong as possible. An alternative representation which will be needed for the Hopeld network solutions is obtained by expressing the SCA problem in terms of a solution vector rather than the solution matrix . Dening

the objective function can be rewritten in the form where is as shown in Fig. 4. The linear constraints (3), which ensure the demand for channels is satised, can be expressed as , where and

subject to (3) and (4) The cost (or proximity factor) tensor described above can be generated according to the recursive relation

. . .

. . .

. . .

. . .

(8) rows. Due to the zero diagonals of , this consisting of representation of the SCA problem is an indenite quadratic form containing many local minima. Consequently, a global minimum cannot be guaranteed by any technique. III. EXISTING TECHNIQUES
FOR

SCA

for for all for all

(5) (6) (7)

The proximity factor tensor for the example compatability matrix (1) is shown in Fig. 3, where the front square of the cuboid is simply the matrix with the diagonal terms overwritten to zero. This is because if and , there should be no penalty incurred (there is effectively only a single call and no interference is possible). The third dimension of the tensor decreases the penalty linearly until the penalty becomes equal to zero. Thus, the effective depth of the tensor is equivalent to the value of the maximum diagonal of . This linear decrease in cost is a means of encouraging less severe violations of the noninterference constraints. If the network demands are such that violation is inevitable, we would prefer

Due to the NP-completeness of the SCA problem, many heuristics have been devised to provide near-optimal solutions to, in particular, CAP1. Finding the minimum number of channels required to obtain an interference-free assignment has become the subject of much research, both algorithmic and theoretical. Naturally, many heuristics arrive at a value for (the number of channels) which is well above the minimum number necessary, and unfortunately, these heuristics provide no indication of how far from optimal their solutions are. The work of Gamst [21] has enabled lower bounds on the minimum number of channels required for an interference-free assignment in a hexagonal network to be calculated, which can be used as a supplement to these heuristics. Many heuristics have been proposed which treat the SCA problem, in particular CAP1, as a graph coloring problem [42]. Consider the graph obtained by representing each call

SMITH AND PALANISWAMI: STATIC AND DYNAMIC CHANNEL ASSIGNMENT USING NEURAL NETWORKS

241

Fig. 4. Matrix of quadratic form

Q for static channel assignment.

degree of interference to be taken into consideration. The use of simulated annealing on CAP2 has also been investigated [9] where several neighborhood transition functions were employed with varying degrees of success. Genetic algorithms have also been recently applied to solve CAP1 [8]. In Section V, we compare the performances of two heuristics (simulated annealing and steepest descent) and a commercial optimization package with the proposed neural techniques on a collection of test problems. The heuristics operate on the SCA problem as formulated in Section II-B. Given an initial feasible solution (satisfying channel requirements), columns of the solution matrix are swapped or moved so that feasibility is always preserved. For the steepest descent heuristic, the current column is swapped with the column which maximizes the decrease in cost to the objective function. For the simulated annealing heuristic, columns are selected at random. Any reduction in cost is automatically accepted as a swap, while swaps generating an increase in cost are accepted if the Boltzmann probability requirements are satised [30]. The cooling schedule for the simulated annealing heuristic is given by

in the network as a vertex, with an edge joining two vertices if there is interference between the corresponding calls. This edge is labeled with the minimum separation distance between the channels assigned to these calls (in cells and ). CAP1 can then be seen as equivalent to the task of assigning positive integers to the vertices of the graph such that if two vertices are connected by an edge, the absolute value of the difference of the integers is at least equal to the edge label, and the maximum integer used is as small as possible. If no ), this reduces adjacent channel constraints exist (i.e., to the classical graph coloring problem, which is known to be NP-complete [22], [23]. Thus, the generalized problem (with adjacent channel constraints) is also NP-complete. Sivarajan et al. [38] describe some algorithms which yield optimal or near-optimal assignments in many cases, and are based on similar exhaustive strategies employed by others [19], [45]. Starting at the top of a list of calls, each call is assigned to the channel with the smallest possible number consistent with previous assignments. These algorithms are noniterative and (where is the number of calls in the network). This are is naturally much faster than the iterative algorithms proposed by Box [3]. Numerous other heuristics have been devised which are based upon graph theoretic approaches [20], [26], [35], [45]. Unfortunately, these approaches do not allow for the solution of problem CAP2 and are therefore not applicable if the demands or constraints are such that no interference-free assignment can be found given the available number of channels. In such a situation it would be appropriate to permit less severe constraints to be violated rst in order that an assignment can be made. Modication of the graph coloring approach to one of a coloring problem of a network [42] has enabled the

where is the temperature during the th Markov chain, is the number of temperature decrements, and is the length of the th Markov chain. The initial and nal temperatures were calculated using the acceptance ratio technique [30] for the entire set of test problems.

242

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 15, NO. 2, FEBRUARY 1997

The commercial optimization package used is GAMS [4] with the nonlinear solver MINOS-5, which is a FORTRANbased system designed to solve linear and nonlinear optimization problems. If the objective function and constraints are both linear, then MINOS-5 implements the primal simplex method of Dantzig [7]. If the objective function is nonlinear (as in the SCA formulation), the reduced gradient method of Wolfe [43] is implemented. While MINOS-5 operates on continuous variables only, we have shown that solving a relaxed version of the SCA using GAMS/MINOS-5 results in integral solutions anyway [41]. This is due to the concavity of the objective function over the constraint plane. In the following sections of this paper, we introduce the proposed neural techniques and demonstrate how they can be used to solve a broad class of 01 combinatorial optimization problems which includes the SCA problem.

choice of parameters suitable for the solution of the SCA problem. A. An Improved Hopeld Network Problems of infeasibility and poor solution quality in the Hopeld network can be essentially eliminated by an appropriate form of energy function and modication of the internal dynamics of the Hopeld network. By expressing all constraints of the problem in a single term, the overall number of terms and parameters in the energy function can be reduced. Consider the general energy function (9) where (10) (11) The rst term of the energy function is the objective function , while the second term measures the deviation of the vector from the constraint plane given by . The advantage of this energy function is that only one penalty parameter, , needs to be selected. If is large enough, then validity of the solution is ensured, since the constraint term will be forced to vanish. Hence, the solution will necessarily lie on the constraint plane. Energy functions of this nature have been suggested by Aiyer [1] and Gee [24]. We now modify the internal dynamics of the Hopeld network to permit temporary increases in this energy function in order to allow escape from local minima. The motivation for this is that the Hopeld network cannot be expected to compete with other hill-climbing heuristics such as simulated annealing while it remains a strict Liapunov descent algorithm. This improved hill-climbing Hopeld network (which we refer to as HCHN) provides a mechanism for escaping local minima by varying the direction of motion of the neurons in such a way that, while descent of the energy function is always permitted, ascent of the energy function is permitted often initially and is less likely as the algorithm proceeds. Clearly, this is similar to the concept of simulated annealing. The HCHN consists, like the Hopeld network [27], [28], of a fully interconnected system of neurons. Neuron has internal state and output level (bounded by 0 and 1). The internal state incorporates a bias current (or negative threshold) denoted by , and the weighted sums of outputs from all other neurons. The weights, which determine the strength of the connections from neuron to , are given by . The relationship between the internal state of a neuron and its output level is determined by an activation function , which is bounded below by 0 and above by 1. The HCHN differs from the Hopeld network in its internal dynamics which permit hill-climbing. The modied differential equation is given by: (12) (13)

IV. NEURAL NETWORK APPROACHES TO THE SCA PROBLEM Neural networks have been used to solve many various problems in the telecommunications industry including adaptive routing in networks, switch scheduling in crossbar switches and multistage interconnection networks, switch control, connection admission control, network reconguration in satellite systems, and routing and scheduling in multihop packet radio networks (see [5] and [44] for a survey of these applications and the neural techniques employed). Hopeld neural networks have also been used to solve the SCA problem [17], [32], [33] with mixed success. Funabiki et al. [17] proposed a parallel algorithm based on the Hopeld network, and solved a set of test problems of varying size in near constant time. Kunz [32], [33], however, implemented the Hopeld network approach exactly as described by Hopeld and Tank [29], using an energy function containing the objective function and individual terms for all constraints of the problem. It is well known that such an approach results in problems of infeasibility, since the many terms of the energy function compete with each other to be minimized. Furthermore, the (strict descent) dynamics of the Hopeld network result in convergence to the rst local minima encountered. Such an approach is therefore unlikely to be competitive with heuristics such as simulated annealing which are able to escape from local minima. In this section, we introduce an improved Hopeld network approach which is able to ensure feasibility of the nal solution and incorporates hill-climbing into its internal dynamics to escape from local minima. We also introduce a new selforganizing neural network approach to optimization. Previous work in this area has been limited to solving the Travelling Salesman Problem and other Euclidean problems [2], [15], [16], [25]. These self-organizing approaches do not generalize to solve other 01 optimization problems such as the SCA problem because they are embedded in the two dimensional plane by their geometry. The new self-organizing neural network (SONN) which we introduce in this section, however, operates only within the framework of the solution matrix and generalizes to solve a variety of problems [40]. For both the improved Hopeld network and the SONN, we discuss the

SMITH AND PALANISWAMI: STATIC AND DYNAMIC CHANNEL ASSIGNMENT USING NEURAL NETWORKS

243

Fig. 5. Graph of k (t) = 1

0 2e0t=

with  = 40.

Furthermore, the usual sigmoidal activation function is replaced by the piecewise linear function

Markov chain (or the number of random walks permitted in multidimensional space) at each point in time is held constant at a value which depends upon the size of the problem. Thus, the modied Hopeld network HCHN allows random increases in energy initially, with such increases becoming less likely as time proceeds, until nally the network tends toward a steepest descent algorithm. Provided the value of the parameter is large enough, the convergence trace will be forced to lie on the constraint plane, resulting in a feasible solution. 1) Parameters for SCA: For the HCHN, the neurons are initialized to small random perturbations around the center of the unit hypercube. The length of the Markov chain is given by , and the value of is set to 40. The effect of the hill-climbing mechanism is tested by comparison with a strict Liapunov descent Hopeld network (which we refer to as HN) which is identical to HCHN with always. B. A Self-Organizing Neural Network SONNs attempt to exploit the principles of neuron organization and the brains apparent ability to nd structure and order from information. The greater biological resemblance of self-organizing approaches, as well as the reduced number of neurons and weights needed to perform optimization tasks, are seen as the major advantages of self-organizing approaches to combinatorial optimization over a Hopeld network. Learning in a self-organizing system occurs due to the processes of competition and inhibition. Upon presentation with an external input, neurons compete with each other to claim the input, and weights are adapted to reect the result of the competition. The idea of channels in a mobile communications network competing to be assigned to calls in the network lends itself well to a self-organizing approach. In this section, we propose a new SON based upon Kohonens self-organizing feature map [31], and modied so that the network is able to solve 01 optimization problems. For the sake of clarity, we discuss the SONN with reference to solving the SCA problem (although the network generalizes to solve a much broader class of problems). We refer the interested reader to our previous work [40] for a more general derivation and discussion of convergence and stability properties. All feasible solutions to the SCA problem, as formulated by (2)(4), lie at vertices of the -dimensional unit hypercube (where ) which also intersect the constraint plane. Since is integer valued for all , the constraint set forms an integral polytope.1 Suppose we allow a continuous approach to such a vertex from within the unit hypercube, starting from a point on the constraint plane and inside the unit hypercube (a feasible, noninteger solution), and gradually moving along the constraint plane to approach a feasible vertex. Let us denote the continuous variable (matrix element) in the interior of the unit hypercube by , so that at the vertices. Essentially, represents the probability that the variable in the th row and th column of the matrix is on. It is the matrix of probabilities (weights), , to which we apply Kohonens principles of self-organization, modifying existing
1 The proof of this involves showing that the matrix unimodular.

(14) within the unit hypercube. The modied so that differential equation means that the direction of the change in any neuron is now controlled by a new parameter , where random and

Fig. 5 shows how the value of changes with time for . If always, then HCHN is simply the original Hopeld network and the states of the network evolve to minimize the energy function (15) provided the weight matrix is symmetric. Now

for within the unit hypercube. If the two energy functions (9) and (15) are made equivalent through selection of and , then

for conned to the constraint plane by a large value of . Thus, steepest descent and ascent of the objective function , on the constraint plane are achieved when respectively. Initially, , and so is randomly selected from the range [ ]. Consequently, the energy value (which is equivalent to the objective cost provided lies , on the constraint plane) will often increase initially. As however, , and so will also approach unity which is needed for strict Lyapunov descent. The length of the

A in (8) is totally

244

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 15, NO. 2, FEBRUARY 1997

Fig. 6. Architecture of SONN.

denitions to enable the process to solve 01 optimization problems. 1) The Network: The architecture of the SONN (shown in Fig. 6) consists of an input layer of nodes, and a linear array of output nodes. The output nodes represent the column indices of the solution matrix (channels), and the input layer represents the cells for the given problem. The weight connecting input node to node of the output layer is given by . A cell in which a channel is required is presented to the network through the input layer, and the nodes of the output layer compete with each other to determine which column of the solution matrix (channel) can facilitate the input vector with least cost. The weights are then adapted to reect this decision using the neighborhood topology. Suppose a channel is required in cell . We present an input vector with a 1 in position (and zero elsewhere) to the network, and for each node of the output layer, we which is the cost to the objective function of calculate assigning channel to cell . Denition 1: The cost potential, , of node for a particular input vector ( ) is (16) Denition 2: The winning node, , of the output layer is the node with minimum cost potential for a particular input vector. That is for all nodes and xed

(as is the case with Kohonens feature map [31]), but is only dened once the cost potential of each node in the output layer has been calculated and ranked for a particular input vector. Therefore, winning nodes and the neighborhood are determined by competition according to the objective function, and the weights are modied according to Kohonens weight adaptation rule within the winning neighborhood. The size of the winning neighborhood is dependent upon which cell is currently being presented to the network. At the end of the Kohonen weight adaptations, the weight matrix has moved in a direction which has most likely taken off the constraint plane and so the solution is infeasible. The next stage of the SONN involves the weights of the other nodes then organizing themselves around the modied weights so that the matrix of network weights remains a feasible solution to the problem at all times. This can be achieved via a Hopeld neural network. Transforming the into a vector (which represents the states weight matrix of the continuous Hopeld network), we perform random and asynchronous updates of (excluding the weights within the winning neighborhood) to minimize the energy function: (17) represents the projection of onto the where constraint plane . The Hopeld network here does not need to employ the hill-climbing dynamic introduced earlier in this paper, since we only need to arrive at a point on the constraint plane and, as such, there is no optimal point. Once the energy function has reached its minimum (so that lies on ), we return to the Kohonen updating stage, presenting another randomly selected channel requirement to the SONN, determining the winning node and its neighbors, and modifying their weights. The entire process is repeated until the network weights stabilize to a feasible 01 solution which is a local minimum of the optimization problem. During convergence, the magnitude of the weight adaptations, and the size of the neighborhoods is gradually decreased. Initially, the size of the neighborhood for each row of [given by ] is large, but is decreased linearly until (the demand for row ) for all . It is worth noting that this self-organizing neural approach is inherently stochastic in nature, since the weight modications made in the SONN are completely dependent upon the order in which the channel requirements are presented. Consequently, the network can be run several times to arrive at different local minima. 2) The Algorithm: 1) Initialize weights of the network as

Denition 3: The neighborhood of the winning node, , is the set of nodes (closest neighbor to farthest neighbor) such that

where is the size of the neighborhood for cell . Thus, the neighborhood of the winning node is not dened spatially, according to the physical architecture of the network

thus giving an initial feasible (noninteger) solution. 2) Randomly select a new channel requirement. Express this as vector (input vector). Find the position (cell) . which is on, i.e., 3) Calculate the potential for each node in the output layer according to (16).

SMITH AND PALANISWAMI: STATIC AND DYNAMIC CHANNEL ASSIGNMENT USING NEURAL NETWORKS

245

4) Choose winning node,

, (by competition) such that

and identify its neighboring nodes

where is the size of the neighborhood for . 5) Update weights in neighborhood of winning node according to

Fig. 7. Twenty-one-cell hexagonal network. TABLE I PROBLEM DESCRIPTIONS FOR HEXAGONAL SCA PROBLEMS

where

(a modied version of Kohonens SOFM updating rule). and are monotonically decreasing and positive functions of time. is a normalized weighting vector used to help the network decide how to break ties for a node. All other weights (not included in neighborhood updating) have . The modied weights are

V. COMPUTATIONAL RESULTS

FOR

SCA

6) The weights will no longer lie on the constraint plane, so we employ a Hopeld neural network to enforce feasibility. is modied around the weight adaptations of the SONN so that . 7) Repeat from Step 2) until all channel requirements in all cells have been selected as input vectors. This is one epoch. Repeat for epochs. Decrease and geometrically. 8) Repeat from Step 2) until . This represents a stable convergence of the weights for a given neighborhood size. Decrease the neighborhood sizes linearly for all . 9) Repeat entire process until . 3) Parameters for SCA: The normalized weighting vector is used to help control oscillations. We use the following formula for this vector for all (18)

The data sets we use in this section to test the performances of the various techniques on the SCA problem can be divided into three classes. The rst class consists of the small example network introduced in Section II-A (which we denote as problem EX1), and a slightly larger extension of this example, described as follows (and denoted by EX2):

and the compatability matrix is given by

The second class of test problems is based upon the 21-cell regular hexagonal network used by Sivarajan et al. [38], as shown in Fig. 7. Two sets of demands for the 21 cells are used and are shown below

This is the same heuristic measure of the degree of difculty of assigning a channel to a cell used by Sivarajan et al. [38]. Each element of is then normalized. The remaining parameters in the SONN are selected as follows:

int

The compatability matrix is generated by considering the rst two rings of cells around a particular cell as interferers. Adjacent channel constraints and co-site constraints are varied, as is the size of the network. The details of these hexagonal test problems (denoted by HEX1 through to HEX4) are shown in Table I, where the number of available channels is the lower bound obtained by applying the lemmata of Gamst [21] to our chosen test problems. The nal set of test problems is derived from the topographical data of an actual 24 21 km area around Helsinki,

246

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 15, NO. 2, FEBRUARY 1997

Fig. 8. Compatability matrix for Kunz data. TABLE II PROBLEM DESCRIPTIONS FOR KUNZ SCA PROBLEMS

Finland, as used by Kunz [32]. Kunz used the computer program GRAND [20] to obtain the expected trafc demand and interference relationships between the 25 regions around the base stations (distributed unequally over the area). Seventythree channels were available. The compatability matrix is shown in Fig. 8, while the demand vector is

The test problems based on this data are obtained by considering only the rst 10 regions (KUNZ1), 15 regions (KUNZ2), 20 regions (KUNZ3), and nally the entire data set (KUNZ4). The number of channels is taken to be a fraction of the 73 channels available for the entire 25 regions, since lower bound techniques are not available for nonhexagonal networks. The exact descriptions of the problems KUNZ1KUNZ4 are shown in Table II. Here, denotes the matrix obtained by taking only the rst rows and columns of . The results presented in Table III compare the performances of GAMS/MINOS-5 (labeled GAMS), and the traditional heuristics of steepest descent (SD) and simulated annealing (SA), with the neural techniques HN (with no hillclimbing), HCHN and the SONN. Each of the techniques (except GAMS/MINOS-5) is run from ten different random starts. Min represents the minimum objective function found during these ten runs, while Av. is the average objective value.

The results for the two small examples, EX1 and EX2, show that all techniques except for GAMS/MINOS-5 are able to nd interference-free assignments (as evidenced by a zero objective value). As expected, the hill-climbing approaches SA and HCHN locate such assignment more frequently than the other techniques. For the hexagonal data set, HCHN obtains the minimum interference solution in every case, although the SA solutions are very competitive with these. HN and the SONN perform better than both the SD heuristic and GAMS/MINOS-5. These relationships are shown in Fig. 9(a). It is interesting to note that, while there are apparently no interference-free assignments possible with the values used in the hexagonal problems (obtained through the of lower bound lemmata of Gamst [21]), the objective values are considerably smaller for problems HEX2 and HEX4. These problems include adjacent channel constraints, enabling tighter lower bounds to be obtained from the lemmata of Gamst. While HCHN out performs HN and SA out performs SD, as should be expected when escape from local minima is permitted, it is also interesting to observe that HN is very competitive with SA for these channel assignment problems. This is especially true in the nal set of results, for problems KUNZ1 through to KUNZ4. Again, HCHN nds the minimum interference assignment in every case. The SONN out performs SD and GAMS considerably. These relationships are depicted in Fig. 9(b). The nal test problem, KUNZ4, is the exact problem used by Kunz [32], where an interferencefree assignment can be found with the available number of channels. Both HCHN and HN were able to obtain such interference-free assignments. HN found six such solutions from the ten random starts, while HCHN located interferencefree solutions in nine of the ten runs, providing evidence that the hill-climbing mechanism successfully enables escape from local minima.

SMITH AND PALANISWAMI: STATIC AND DYNAMIC CHANNEL ASSIGNMENT USING NEURAL NETWORKS

247

RESULTS

OF

TABLE III SCA TEST PROBLEMS FOR GAMS/MINOS-5, SD, SA, HN, HCHN,

AND

SONN

(a)

(b)

Fig. 9. Comparison of techniques on (a) SCA hexagonal test problems and (b) SCA Kunz test problems.

Furthermore, every run of all the neural techniques tested resulted in a feasible solution, which is a claim which cannot be made with a standard Hopeld network implementation. It should be mentioned here that no CPU times have been provided for these solutions. This omission is intentional, since the neural networks are designed to be run on suitable hardware, not simulated on a digital computer. Our simulations have been used in this paper merely to provide an indication of the kind of solution quality which can be expected from the techniques, and as such, we are not interested in simulation times. VI. EXTENSIONS TO DYNAMIC CHANNEL ASSIGNMENT While solutions to the static channel assignment problem are useful for planning a cellular mobile communications network, and for its operation in the immediate future thereafter, it will become increasingly more difcult to nd satisfactory assignments as the demands increase beyond some point. The ability to dynamically alter the channel assignments is imperative as demands increase, if efciency of the frequency spectrum is to continue to be a goal of the optimization process. There have been several approaches proposed for solving the DCA problem. One method is to maximize the total number of available channels over the network when assigning each call, and if no channel can be found the re-

use area is searched for a swap. If such a swap exists, this will free a previously busy channel in the re-use area which can now be assigned to the call. The method of searching for channels to swap within the re-use area is known as a limited rearrangement strategy. Other approaches have attempted to minimize the probability of a handover. This occurs when a call cannot be serviced by the channel to which it is assigned for the entire duration of the call and needs to be handed over to another channel while the call is still in progress. It is important to avoid handovers during peak times since there may be no available channel on which to continue the call which will then be dropped. DuqueAnt n et al. [9] have suggested the use of simulated annealing o in a running system to minimize the total probability of a handover. Due to the increasing demands, the only practical way to handle the network is by use of micro-cells, where the diameter of the cells is decreased to about 100 or 200 m in urban surroundings, or to even smaller values in crowded areas. Duque-Ant n et al., observed, however, that the amount o of interference induced by the simultaneous use of a channel in two cells is not known in practice in this environment, since the details of the trafc cannot be as precisely known once micro-cells are used. Consequently, they have devoted much research to the learning of the compatability matrix in an adaptive environment [10], [11] by statistically correlating carrier and interference power levels with call activities.

248

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 15, NO. 2, FEBRUARY 1997

Graph theoretic approaches have also been used for DCA [37], including the maximum packing strategy of Everitt et al. [13], [14]. This is an idealization of a dynamic channel assignment algorithm, which assumes that a new call will be blocked only if there is no possible reallocation of channels to calls (including reallocating calls in progress) which would allow the call to be carried. Given the number of calls in progress in all cells at any time, the minimum number of channels required to carry these calls can be determined using a graph coloring algorithm. The maximum packing strategy then assumes that this minimum number is the actual number of channels which are in use, while partial rearrangement of calls in progress can be made using existing algorithms [6], [12], [36]. Many hybrid schemes have also been proposed in which combinations of static and dynamic channel assignment are used in various parts of the network. Nakano et al. [34] apply a DCA algorithm to the city part of a service area, while retaining a SCA scheme for the suburban areas. The advantage of this hybrid approach is that efciency of channel usage can be improved by using DCA in the busier areas, while SCA elsewhere means that expensive equipment (used for DCA) can be spared. Using the SONN for DCA Although this idea is still in its preliminary stages, we have included it here to demonstrate how (conceptually) a self-organizing system can be used to dynamically rearrange channel assignments. The SONN appears to be suited (at least conceptually) to DCA since it is an inherently adaptive system. Presenting a new input vector can be seen as analogous to the arrival of a new call in the network, altering the demand vector and the probability distribution of calls simultaneously. Suppose we start with the SCA solution at the point where a new call cannot be assigned to a channel without the need for rearrangement of the existing assignments. The vector of demands is now a function of time, so that constraint satisfaction during epoch is enforced by . Each epoch represents the arrival of single or multiple new calls to the network and, during that epoch, input vectors (denoting the cell in which a call is made) are presented with a frequency of occurrence determined by the probability distribution of the calls at that time. Since feasibility is always restored during the second stage of the SONN, any rearrangement of the existing calls to enable a new call to be accepted is automatic. If no such rearrangement exists, the SONN may be either unable to converge back onto the constraint plane, or a feasible rearrangement may be found but the interference levels could be greater than those considered acceptable. In either case the call can be blocked and the previous state of the system reinstated. The size of the neighborhood function is also adaptive and should be made equivalent to the demand vector during each epoch. It can be seen that the SONN naturally implements a maximum packing strategy, and that the rearrangement of existing assignments is automatically implemented every time a new call arrives in the system. The soft objective approach enables a rearrangement to be found which may slightly violate

the noninterference constraints, where other techniques would conclude that no rearrangement exists and the call would be blocked. Such an approach may be preferable considering the highly competitive nature of the telecommunications industry. Of course, implementation of the SONN for DCA in a realtime adaptive system would need to be extremely fast. More precisely, the time taken to iterate one epoch would need to be less than the arrival rate of new calls to the system. Clearly, simulation of the SONN on a digital computer cannot achieve the required speed, and we rely upon the promise of hardware implementation to convert this idea into a real-time adaptive approach to DCA. VII. CONCLUSION In this paper, we have considered the problem of assigning channels to calls in a cellular mobile communications network. The problem is of great importance in the telecommunications industry, nding application not just to cellular networks, but in satellite and other systems where the available frequency spectrum is a limited resource. While many researchers have considered the problem of nding the minimum number of channels required to obtain an interference-free assignment, it is immensely more useful in practical situations, where the number of available channels is substantially less than this minimum number, to nd the assignment which minimizes the severity of any interferences given the number of available channels. This is the approach we have taken in this paper. The SCA problem has been formulated as a generalized quadratic assignment problem with the interference constraints in the objective function, and the demands treated as hard constraints. We have proposed two neural network approaches to solving the SCA problem. The rst of these is an improved Hopeld network capable of hill-climbing to escape from local minima and improve solution quality. The second approach is the use of Kohonens concepts of self-organization, which we have signicantly modied to enable an optimization process. In previous work, we have successfully applied both of these neural network approaches to solve a variety of optimization problems of practical signicance, including sequencing and location-allocation [39][41]. This paper has shown the potential of neural techniques for solving problems of interest to the telecommunications industry. Three different classes of SCA problems have been considered, from which it has been shown that the neural techniques perform extremely well when compared to a commercial optimization package, and steepest descent and simulated annealing heuristics. Our SONN may also be well suited to solving the DCA problem, due to its inherently adaptive nature. Implementation of the SONN for this purpose will be the subject of future research. REFERENCES
[1] S. V. B. Aiyer, Solving combinatorial optimization problems using neural networks, Cambridge Univ. Eng. Dept., Cambridge, U.K., Tech. Rep. CUED/F-INFENG/TR 89, 1991. [2] B. Ang niol, G. De La Croix, and J.-Y. Le Texier, Self organizing e feature maps and the travelling salesman problem, Neural Networks, vol. 1, pp. 289293, 1988. [3] F. Box, A heuristic technique for assigning frequencies to mobile radio nets, IEEE Trans. Veh. Technol., vol. VT-27, no. 2, pp. 5764, 1978.

SMITH AND PALANISWAMI: STATIC AND DYNAMIC CHANNEL ASSIGNMENT USING NEURAL NETWORKS

249

[4] A. Brooke, D. Kendrick, and A. Meeraus, GAMSA Users Guide. CA: Scientic Press, 1990. [5] B. S. Cooper, Selected applications of neural networks in telecommunication systemsA review, Austr. Telecommun. Res., vol. 28, no. 2, 1994. [6] D. C. Cox and D. O. Reudink, Increasing channel occupancy in large scale mobile radio systems: Dynamic channel reassignment, IEEE Trans. Veh. Technol., vol. VT-22, pp. 218222, 1973. [7] G. B. Dantzig, Linear Programming and Extensions. Princeton, NJ: Princeton University Press, 1963. [8] R. Dorne and J.-K. Hao, An evolutionary approach for frequency assignment in cellular radio networks, in Proc. IEEE Int. Conf. Evolut. Comput., 1995, pp. 539544. [9] M. Duque-Ant n, D. Kunz, and B. R ber, Static and dynamic channel o u assignment using simulated annealing, in Neural Networks in Telecommunications, B. Yuhas and N. Ansari, Eds. Boston, MA: Kluwer, 1994. [10] M. Duque-Ant n, P. Koch, D. Kunz, B. R ber, and M. Ullrich, An o u adaptive method to learn the compatability matrix for microcellular systems, in Proc. Veh. Technol. Conf., Stockholm, Sweden, 1994. , Learning the compatability matrix for adaptive resource man[11] agement in cellular radio networks, Eur. Trans. Telecommun., vol. 6, no. 6, 1995. [12] S. M. Elnoubi, R. Singh, and S. C. Gupta, A new frequency channel assignment algorithm in high capacity mobile communications systems, IEEE Trans. Veh. Technol., vol. 21, pp. 125131, 1982. [13] D. E. Everitt and N. W. Macfadyen, Analysis of microcellular mobile radiotelephone systems with loss, Br. Telecom Tech. J., vol. 1, no. 2, pp. 3745, 1983. [14] D. Everitt and D. Manseld, Performance analysis of cellular mobile communications systems with dynamic channel assignment, IEEE J. Select. Areas Commun., vol. 7, no. 8, pp. 11721180, 1989. [15] F. Favata and R. Walker, A study of the application of Kohonen-type neural networks to the travelling salesman problem, Biol. Cybern., vol. 64, pp. 463468, 1991. [16] J. C. Fort, Solving a combinatorial problem via self-organizing process: An application of the Kohonen algorithm to the traveling salesman problem, Biol. Cybern., vol. 59, pp. 3340, 1988. [17] N. Funabiki and Y. Takefuji, A neural network parallel algorithm for channel assignment problems in cellular radio networks, IEEE Trans. Veh. Technol., vol. 41, no. 4, pp. 430437, 1992. [18] A. Gamst, Homogeneous distribution of frequencies in a regular hexagonal cell system, IEEE Trans. Veh. Technol., vol. 31, no. 3, pp. 132144, 1982. [19] A. Gamst and W. Rave, On frequency assignment in mobile automatic telephone systems, in Proc. GLOBECOM, 1982, pp. 309315. [20] A. Gamst, R. Beck, R. Simon, and E.-G. Zinn, An integrated approach to cellular radio network planning, in Proc. 35th IEEE Veh. Technol. Conf., 1985, pp. 2125. [21] A. Gamst, Some lower bounds for a class of frequency assignment problems, IEEE Trans. Veh. Technol., vol. 35, no. 1, pp. 814, 1986. [22] M. R. Garey and D. S. Johnson, The complexity of near-optimal graph coloring, J. ACM, vol. 23, pp. 4349, 1976. , Computers and Intractability. New York: W. H. Freeman, [23] 1979. [24] R. S. Garnkel and G. L. Nemhauser, Integer Programming. New York: Wiley, 1972. [25] A. H. Gee, Problem solving with optimization networks, Ph.D. dissertation, Queens College, Cambridge, U.K., 1993. [26] M. Goldstein, Self-organizing feature maps for the multiple traveling salesman problem (MTSP), in Proc. IEEE Int. Conf. Neural Networks, Paris, France, 1990, pp. 258261. [27] W. K. Hale, Frequency assignment: Theory and applications, Proc. IEEE, vol. 68, pp. 14971514, 1980. [28] J. J. Hopeld, Neural networks and physical systems with emergent collective computational abilities, Proc. Nat. Acad. Sci., pp. 25542558, 1982. [29] , Neurons with graded response have collective computational properties like those of two-state neurons, Proc. Nat. Acad. Sci., pp. 30883092. [30] J. J. Hopeld and D. W. Tank, Neural computation of decisions in optimization problems, Biol. Cybern., vol. 52, pp. 141152, 1985. [31] S. Kirkpatrick, C. Gelatt, and M. Vecchi, Optimization by simulated annealing, Science, vol. 220, pp. 671680, 1983. [32] T. Kohonen, Self-organized formation of topologically correct feature maps, Biol. Cybern., vol. 43, pp. 5969, 1982. [33] D. Kunz, Channel assignment for cellular radio using neural networks, IEEE Trans. Veh. Technol., vol. 40, no. 1, pp. 188193, 1991. [34] , Suboptimal solutions obtained by the Hopeld-Tank neural network algorithm, Biol. Cybern., vol. 65, pp. 129133, 1991.

[35] K. Nakano, M. Yokono, M. Sengoku, Y. Yamaguchi, S. Shinoda, S. Motooka, and T. Abe, An application of dynamic channel assignment to a part of a service area of a cellular mobile communication system, IEICE Trans. Fund. Electr., vol. 75, no. 3, pp. 369379, 1992. [36] R. J. Pennotti, Channel assignment in cellular mobile telecommunications systems, M.Sc. thesis, Polytechnic Institute of New York, Brooklyn, 1976. [37] M. Sengoku, M. Kurata, and Y. Kajitani, Rearrangement of switching networks and its applications to a mobile radio communications system, in Proc. IEEE Int. Conf. Circuits Comput., 1980, pp. 402405. [38] M. Sengoku, Telephone trafc in a mobile radio communication system using dynamic frequency assignments, IEEE Trans. Veh. Technol., vol. 29, no. 2, pp. 270278, 1980. [39] K. N. Sivarajan and R. J. McEliece, Channel assignment in cellular radio, in Proc. IEEE Veh. Technol. Conf., 1989, pp. 846850. [40] K. Smith, M. Palaniswami, and M. Krishnamoorthy, Traditional heuristic versus Hopeld neural network approaches to a car sequencing problem, Eur. J. Oper. Research, vol. 93, no. 2, pp. 300316, 1996. , A hybrid neural approach to combinatorial optimiza[41] tion,Comput. Oper. Res., vol. 23, no. 6, pp. 597610, 1996. [42] K. Smith, Solving combinatorial optimization problems using neural networks, Ph.D. dissertation, University of Melbourne, Australia, 1996. [43] H. Tamura, M. Sengoku, S. Shinoda, and T. Abe, Channel assignment problem in a cellular mobile system and a new coloring problem of networks, IEICE Trans., vol. E-74, no. 10, pp. 29832989, 1991. [44] P. Wolfe, The reduced gradient method, unpublished manuscript, RAND Corp., 1962. [45] B. Yuhas and N. Ansari, Eds., Neural Networks in Telecommunications. Boston, MA: Kluwer, 1994. [46] J. A. Zoellner and C. A. Beall, A breakthrough in spectrum conserving frequency assignment technology, IEEE Trans. Electromagn. Compat., vol. 19, pp. 313319, 1977.

Kate Smith (M96) received the Bachelor of Science (Honors) and Ph.D. degrees from the University of Melbourne, Victoria, Australia, in 1992 and 1996, respectively. She is a Lecturer in the Department of Business Systems, Monash University, Australia. Her research interests are in neural networks, combinatorial optimization, communications systems, and applications to business and industry.

Marimuthu Palaniswami (S84M85SM94) received the Master of Engineering Science degree from the University of Melbourne, Victoria, Australia, in 1984, and the Ph.D. from the University of Newcastle, Australia, in 1987. Since 1987, he has been with the University of Melbourne, Victoria, Australia, where he is an Associate Professor. His research interests are in the elds of neural networks, image processing, communication systems, and intelligent control. He has published more than 100 conference and journal papers in these topics. He is an Associate Editor of the IEEE TRANSACTIONS ON NEURAL NETWORKS and is on the Editorial Board of two other journals. He was a Technical Program Co-Chair for the IEEE International Conference on Neural Networks in 1995. He has served on the Program Committees for a number of conferences including the IEEE Conference on Neural Networks, 1994, 1995, 1996, the IEEE Workshop on Emerging Technologies and Factory Automation, and the Third, Fourth, and Fifth Australian Conference on Neural Networks. He was also the Technical Program Chair for the rst IEEE AustraliaNew Zealand Conference on Intelligent Information Processing Systems in 1993. He has active collaborative links with a number of international universities including the Florida International University, the Indian Institute of Science, the Nanyang Technological University, and the Communications Research Laboratory, Japan. Dr. Palaniswami received the Foreign Specialist Award from the Ministry of Education, Japan.