A Neural Fuzzy Resource Manager for Hierarchical
Cellular Systems Supporting Multimedia Services
Kuen-Rong Lo, Chung-Ju Chang, Senior Member, IEEE, and C. Bernard Shung
Abstract—Using intelligent techniques to perform radio resource management is an effective method. This paper pro-poses neural fuzzy control for radio resource management in hierarchical cellular systems supporting multimedia services. A neural fuzzy resource manager (NFRM) is designed, which mainly contains a neural fuzzy channel allocation processor (NFCAP), and NFCAP is in a two-layer architecture: a fuzzy cell selector (FCS) in the first layer and a neural fuzzy call-admission and rate controller (NFCRC) in the second layer. The FCS chooses not only the handoff failure probabilities and the resource availabilities in both microcell and macrocell but also the mobility of user as input linguistic variables. The NFCRC takes the handoff failure probability and the resource availability of the selected cell as input variables to perform call admission control and rate control for the call. Simulation results show that NFRM can always guarantee the quality of service (QoS) requirement of handoff failure probability for all traffic loads. Also, NFRM improves the system utilization by 31.1% while increasing the handoff rate by 2% over the overflow channel allocation (OCA) scheme [3]; it enhances the system utilization by 6.3% and 1.4%, and still reduces the handoff rate by 14.9% and 6.8%, as compared to the combined channel allocation (CCA) scheme [20], [21] and fuzzy channel allocation control (FCAC) scheme [9], respectively, under a predefined QoS constraint.
Index Terms—Call admission, hierarchical cellular systems, macrocell, microcell, neural fuzzy, resource management.
I. INTRODUCTION
T
HE future mobile communication system will provide not only voice and low-speed data services but also high-speed multimedia services [1], [2]. A way to provide a wide variety of services is to flexibly aggregate multiple channels (time-slot or spreading code), without changing the spectrum division, mod-ulation, and burst structure. A mobile station (MS) specifies the required capacity and desired capacity in the setup request or handoff request message. The required and desired capacities characterize the service requirements of an application, or the patience of a user. Based on the availability of resources in a cell and the quality-of-service (QoS) requirement, the network gives the MS a number of channels between the required capacity and desired capacity. On the other hand, a hierarchical cellular struc-ture, which contains overlaid microcells for high-teletraffic areaManuscript received May 17, 2001; revised December 31, 2002. This work was supported by the National Science Council, Taiwan, R.O.C., under Contract NSC 87-2218-E009-047 and NSC 88-2213-E009-127.
K.-R. Lo is with Telecommunication Laboratories, Chunghwa Telecom Co., Ltd., Chung-Li, Taiwan, R.O.C.
C.-J. Chang is with National Chiao Tung University, 300 Taiwan, R.O.C. (e-mail: cjchang@cc.nctu.edu.tw).
C. B. Shung is with Broadcom Corp., San Jose, CA 95134 USA. Digital Object Identifier 10.1109/TVT.2003.816002
and overlaying macrocells for low-teletraffic region, has merits of enhancing system capacity and improving coverage [3]–[6]. Such a wireless network must be adaptive and robust to support resource demands.
Nowadays, the intelligent techniques have been widely applied to nonlinear, time-varying, and complicated problems that were challenging using conventional algorithmic methods. These techniques such as fuzzy logics, neural networks, and neural fuzzy networks have been shown to outperform algorithmic methods. The advantages of intelligent techniques are numerous, most notably learning from experience and the scalability, adaptability, and ability to extract rules without the need for detailed or precise mathematical modeling [7]–[16].
In this paper, we propose a neural fuzzy resource manager (NFRM) for hierarchical cellular systems providing multimedia services. The NFRM utilizes the learning capability of the neural network to reduce the decision error of these conven-tional channel assignment schemes resulting from modeling, approximation, and unpredictable statistical fluctuations of the system. It also employs the control rule structure of fuzzy logic, which absorbs benefits of those conventional channel assignment schemes, to provide robust operation and to prevent operating errors due to the learning of incorrect training data. The NFRM contains a neural fuzzy channel allocation processor (NFCAP), a resource estimator, a performance evaluator, and base-station interface modules. NFCAP is a two-layer neural fuzzy logic controller that consists of a fuzzy cell selector (FCS) in the first layer and a neural fuzzy call-ad-mission and rate controller (NFCRC) in the second layer. The FCS considers the handoff failure probability, the resource availability in both macrocell and microcell, and the mobility of users as input linguistic variables, and applies the max-min interference method to determine which cell, macrocell or microcell, to serve the call request; FCS intends to enhance the channel utilization by balancing utilization between macrocell and microcells. The NFCRC takes the handoff failure proba-bility and the resource availaproba-bility of the selected cell as input variables; NFCRC intends to guarantee the QoS and provides an appropriate rate for users. Simulation results show that NFRM can guarantee the QoS requirement of handoff failure probability for all traffic loads. NFRM improves the system utilization by 31.1% while increasing the handoff rate by 2% over the overflow channel allocation (OCA) scheme proposed in [3]; and it enhances the system utilization by 6.3% and 1.4%, and still reduces the handoff rate by 14.9% and 6.8%, as compared to the combined channel allocation (CCA) and fuzzy channel allocation control (FCAC) scheme proposed in [20], [21] and [9], respectively, under a defined QoS constraint. 0018-9545/03$17.00 © 2003 IEEE
LO et al.: NEURAL FUZZY RESOURCE MANAGER FOR HIERARCHICAL CELLULAR SYSTEMS 1197
Fig. 1. The NFRM for hierarchical cellular systems.
The rest of this paper is organized as follows. Section II presents the functions of NFRM. Section III gives the design of NFCAP. Section IV shows simulation results and discussions. Conclusions are given in Section V.
II. NEURALFUZZYRESOURCEMANAGER(NFRM) Fig. 1 shows the functional block diagram of NFRM for hier-archical cellular systems supporting multimedia services, where the hierarchical cellular system contains a large geographical region tessellated by cells, referred to as macrocells, each of which overlays several microcells. The overlaying macrocell is denoted by cell 0 and its overlaid microcells are denoted by cell , . The NFRM contains functional blocks such as base-station interface module (BIM), resource estimator, per-formance evaluator, and neural fuzzy channel allocation pro-cessor (NFCAP). BIM is to interface with macrocell or micro-cell base stations. It is installed in the base-station controller (BSC) or mobile switching center (MSC). Note that for sim-plicity, NFRM is drawn to do the resource management for only one macrocell here.
Cell in the macrocell is equipped with a pool of inde-pendent communication channels, . Assume all the channels are shared by new calls and handoff calls. The new call requests generated in each MS is modeled as a Poisson process with mean rate , in which the arrival rate of the new voice calls is and the arrival rate of the new data calls is , . The call durations for the two streams are assumed to be exponentially distributed with aver-ages equal to 1 and 1 . We assume that all the data (voice) call requests have identical required capacity and desired ca-pacity, denoted by and ( and ), respectively.
A. Base-Station Interface Modules (BIM)
Assume that each BIM provides complete-partitioning buffers for queueing new and handoff calls, which are originated in the corresponding cell and temporarily have no free channel to use. In the BIM for cell 0 are a new-call buffer with capacity for new calls originating in the macrocell-only region, a handoff-call buffer with capacity for handoff calls from adjacent macrocells, an overflowed handoff- call buffer with capacity for overflowed handoff calls from overlaid mi-crocells, and an overflowed new-call buffer with capacity for overflowed new calls from overlaid microcells. In the BIM for cell (BIM ), , there are a new-call buffer with capacity for new call originations, an underflowed handoff-call buffer with capacity for underflowed handoff calls from the overlaying macrocell, and a handoff-call buffer with capacity for handoff calls from adjacent microcells. No buffer is provided for the reversible handoff calls. Reneging of new calls and dropping of handoff calls are considered be-cause of new calls’ impatience and handoff calls’ moving out the handoff area. The patience times are exponentially distributed.
Whenever BIM receives a call request, , it sends the necessary calling information to the resource estimator, the performance evaluator, and the NFCAP. The calling information can distinctly indicate from which cell and in what type the call is originated. The type of call is defined as: denotes a new call originating in macrocell-only region; denotes a handoff call from adjacent macrocell to macrocell-only region; denotes a handoff call from microcell to macrocell-only region; denotes new call originating in microcell; denotes handoff call from adjacent macrocell to an overlaid mi-crocell; denotes handoff call from microcell to microcell; and denotes reversible handoff. Note that the
macro-TABLE I
THECALCULATION OFAVAILABLERESOURCE
cell-only region is the area inside macrocell 0 but outside all microcells. The first three types of calls are to use channels in macrocell, while other types of calls can use channels either in macrocell or in microcell.
B. Resource Estimator
The resource estimator calculates the available resources in macrocell 0 and microcell when it receives calling information of type- call from BIM at time instant , which are denoted by and , respectively. The resource estimator knows
system parameters of , , , , , and , ,
, , , and it obtains and by
formulas shown in Table I.
In Table I, is the number of occupied channels in
at time ; ( , , ) is the number
of waiting calls in the new-call buffer (handoff-call buffer, overflowed handoff-call buffer, overflowed new-call buffer) of BIM , at time ; and ( , ) is the number of waiting calls in the new-call buffer (underflowed handoff-call buffer, handoff-call buffer) of BIM at time , . C. Performance Evaluator
The performance evaluator is to calculate the handoff failure probability for NFCAP. The handoff failure probability in macrocell (microcells) at time , denoted by , is defined as
(1) where is the number of blocked handoff calls in macrocell 0 (microcell ) at time ; is the number of dropped handoff calls in macrocell 0 (microcell ) at time ; and is the number of handoff calls in macrocell 0 (microcell ) at time .
D. Neural Fuzzy Channel Allocation Processor (NFCAP) NFCAP performs the channel allocation using neural fuzzy logic control to attain QoS guaranteed, high channel utilization, and good user satisfaction. In the neural fuzzy logic control, a reinforcement learning is designed to adjust the mean and the variance of the membership functions to cope with the input traffic fluctuation. The detailed design of NFCAP is described in the next section.
III. DESIGN OFNFCAP
NFCAP contains two functional blocks: FCS in the first layer and NFCRC in the second layer, as shown in Fig. 2.
A. Fuzzy Cell Selector (FCS)
NFCAP chooses five input linguistic variables for FCS: available resources in macrocell 0 and in microcell
, handoff failure probabilities in macrocell 0
and in microcell , and mobile speed , and has one output linguistic variable for FCS: the selection of macrocell or microcell . The available resource of cells can indicate the remaining capacity, the handoff failure probability can show the QoS, and the mobile speed can implicate the handoff rate. Term sets for both and are
{More Enough, Slightly Enough, Not
Enough} , term sets for both
and are {Low,
Medium, High} , and the term set for is
Slow, Fast . A trapezoidal function is chosen to implement the membership function, which is given by
for for for otherwise
(2) where in is the left (right) edge of the trapezoidal function and is the left (right) width of the trapezoidal function.
Denote , ,
and as the membership functions
for , , and in ,
respec-tively, and define , , ,
, , and as (3) (4) (5) (6) (7) (8) The maximum possible “More Enough” value of available resource would be the sum of buffer size and allocation channels, would be a safety margin of available resource in macrocell (microcells) in QoS require-ment and traffic fluctuation, would be set to be a fraction of available resource in macrocell (microcells),
, and and
are provided to tolerate the change of traffic in macrocell (microcells).
Denote , and ,
, and to be the membership functions
LO et al.: NEURAL FUZZY RESOURCE MANAGER FOR HIERARCHICAL CELLULAR SYSTEMS 1199
Fig. 2. The block diagram of NFCAP.
TABLE II
THEINFERENCERULES FOR THEOVERLAYREGION
define , , , , , and as (9) (10) (11) (12) (13) (14) would be set to be provided to guarantee the QoS requirement in macrocell (microcells), would be set to be a safety margin of the handoff failure probability in QoS requirement in macrocell (microcells),
, and and
are provided to tolerate the dynamic behavior of the handoff failure probability in macrocell (microcells).
The membership functions for terms and in , denoted
by and , are given by
(15) (16)
where would be a fraction of slow (fast) speed of mobile user, is provided to tolerate the change of slow (fast) speed, and would be the fastest speed.
There are different call types in hierarchical cellular sys-tems. For calls that can use only macrocell channels, FCS has to choose the macrocell, and send and to NFCRC. For calls that could use channels either in macrocell or microcell, FCS determines the serving cell according to
input linguistic variables of , , , ,
and . The output linguistic variable if the macro-cell is assigned and if the microcell is allocated. . The fuzzy rule base with dimension is shown in Table II, where denotes the number of terms
in .
The design idea of the fuzzy rule structure listed in Table II is described as follows. If the available resource in macrocell is larger than that in microcell , the call would be directed toward the macrocell , and vice versa . If the available resources in both macrocell and microcell have the same fuzzy terms in the premises of the fuzzy rule, the call will be directed to a cell with low handoff failure probability ( or ), instead of the one with high handoff failure probability. If the available resource and
handoff failure probability in both macrocell and microcell have the same fuzzy terms in the premises of the fuzzy rule, then the call is to be biased toward macrocell if the speed is fast, for lessening frequent handoff, and vice versa. Note that the symbol “-” in the table represents no impact on the output of the fuzzy cell selector.
Membership functions for and in are defined as
(17) (18) We adopt the max-min inference method and apply the center-of-area defuzzification method for output variable [9], which are not further described here. There are and output to NFCRC determined by : if the call is with channels
in the macrocell , and ;
otherwise, the call is with channels in the microcell ,
and , .
B. Neural Fuzzy Call-Admission and Rate Controller (NFCRC)
The NFCRC takes the handoff failure probability and available resource as input linguistic variables. The handoff failure probability shows the QoS, and the available resource implicates the traffic load intensity. This is a feedback control system that the handoff failure probability acts as a QoS index feedback to indicate how effectively the NFCRC is managing the radio resource.
We adopt a five-layer neural fuzzy controller to design the NFCRC. The best structure of NFCRC can be obtained via structure learning, which measures the degree of fuzzy similarity and decides the size of the fuzzy partition of the linguistic [17], [18]. Usually, a hybrid learning algorithm is applied to construct the NFCRC. The algorithm is a two-phase learning scheme. In phase I, a self-organized learning scheme is used to construct the presence of rules and to locate the initial membership functions; in phase II, a reinforcement learning scheme is used to optimally adjust the membership functions for desired outputs. To initiate the learning process, the size of the term set for each input/output linguistic variable, the fuzzy control rules, and training data must be provided. In the self-organized training phase, the initial structure of the controller could be constructed via Kohonen’s feature-maps algorithm and the N nearest neighbors scheme [19] to provide a rough estimate of the structure, if the controller is not provided with an initial knowledge base. However, in this paper, we construct an initial form of the controller based on the domain knowledge obtained from the fuzzy channel allocation control scheme proposed in [9]. Only a slight modification for the structure is needed in the self-organized training phase.
The rule structure for NFCRC is shown in Table III. The de-sign strategy in Table III is that if the handoff failure proba-bility is Low (L) or Medium (M) and the available resource is not Not Enough (NE), the call would have a chance to enter the system; if the available resource is Not Enough (NE), the call would be Rejected (R) or Weakly Rejected (WR); if the handoff
TABLE III
THEINFERENCERULES FORNFCRC
failure probability is High (H) and the available resource is More Enough (ME), the call would be Weakly Accepted (WA) for in-creasing channel utilization. If the handoff failure probability is High (H) or available resource is Not Enough (NE), the call would be allocated Basic Rate (BR); if the handoff failure prob-ability is Medium (M) and the available resource is Slightly Enough (SE), the call would be allocated Low Medium Rate (LM); if the handoff failure probability is Medium (M) (Low (L)) and the available resource is More Enough (ME) (Slightly Enough (SE)), the call would be allocated High Medium Rate (HM); if the handoff failure probability is Low (L) and the avail-able resource is More Enough (ME), the call would be allocated High Rate (HR).
The connectionist structure of the NFCRC is constructed in Fig. 3. The NFCRC has the nodes in layer one as input linguistic nodes. It has two pairs of nodes in layer five, where each pair of output nodes has two kinds of linguistic nodes. One is for feeding training data (desired output) into the net and the other is for pumping decision signals (actual output) out of the net. The nodes in layer two and layer four are term nodes, which act as membership functions of the respective linguistic variables. The nodes in layer three are rule nodes; each node represents one fuzzy rule and all nodes form a fuzzy rule base. The links in layer three and layer four function as an inference engine; layer-three links define preconditions of the rule nodes and layer-four links define consequences of the rule nodes. The links in layer two and layer five are fully connected between the linguistic nodes and their corresponding term nodes.
NFCRC has a net input function and an activation output function for node in layer , where de-notes the input to node in layer from node in layer ( 1). The layers are described in the following.
Layer 1: In this layer, there are two input nodes with the respective input linguistic variables and .
Define
(19)
where and .
Layer 2: The nodes in this layer are used as the fuzzifier. The term set used to describe the handoff failure proba-bility is defined as {Low (L), Medium (M), High (H)}. And the term set for the available resource is de-fined as {More Enough (ME), Slightly Enough (SE), Not Enough (NE)}. Thus we have six nodes in this
LO et al.: NEURAL FUZZY RESOURCE MANAGER FOR HIERARCHICAL CELLULAR SYSTEMS 1201
Fig. 3. The structure of the NFCRC controller.
layer. Each node performs a bell-shaped function defined as
and
(20)
where , , and and
are the mean and the standard deviation of the th term of the input linguistic variable from node in the input layer,
respectively, if and if .
Layer 3: The links perform precondition matching of fuzzy control rules. According to fuzzy set theory, the fuzzy rule base forms a fuzzy set with dimensions . Thus, there are nine rule nodes in this layer. And each rule node performs the fuzzyAND operation defined as
and
(21)
where and all that are precondition
nodes of the th rule.
Layer 4: There are two groups of output in this layer: one group for the output of admission control and the other group for the output rate control . Nodes in this layer have two operating modes: down-up and up-down. In the down-up operating mode, the links perform consequence matching of fuzzy control rules. In order to provide a soft admission decision, the term set of the output linguistic variable is defined as {Reject (R), Weakly
Reject (WR), Weakly Accept (WA), Accept (A)}. Simi-larly, the term set of the output linguistic variable is defined as Basic Rate (BR), Low Medium Rate (LM), High Medium Rate (HM), High Rate (HR)}. Thus, there are eight nodes in this layer. And each node performs a fuzzyORoperation, which integrates the fired strength of rules that have the same consequence and is defined as
and
(22)
where and all that have the same
consequence of the th term in the term set of and . The up-down operating mode is used during learning pe-riods, which will be described later.
Layer 5: There are two pairs of nodes in this layer. One node in each pair performs the down-up operation for the decision signals and . The node and its links act as the defuzzifier. The function used to simulate a center-of-area defuzzification method for signal is approxi-mated by
and
(23)
where , is the decision threshold, and if
Clearly, , and a new connection will be accepted only if . Similarly, the signal is also to simulate a center-of-area defuzzification method approximated by
and
(25)
where ; is the number of desired channels for a call; denotes the voice call; and denotes the data call. Clearly, and a new connection is assigned to use a number of channels. The other node performs the up-down operation during the learning period.
The procedure to locate the mean of the th mem-bership function for linguistic variable , , is described below, given a set of training data for , . It employs the statistical clustering technique of Kohonen’s feature-maps algorithm [19]. This is the ini-tial definitions of membership functions required to drive the reinforcement learning algorithm.
Obtain by using Kohonen’s feature-maps algorithm
Step 1: Set initial values of for all
membership functions, , such
that
Set an initial learning rate .
Step 2: Set .
Step 3: Present training data and
com-pute the distance , .
Step 4: Determine the th membership func-tion that has the minimum distance
. Update by
Step 5: If , , Goto Step 3 ELSE
Decrease and Goto Step 2.
EndIf
The above procedure will stop until . The determination of which is minimum at Step 4 can be quickly accomplished in constant time via a winner-take-all circuit [19]. The adaptive algorithm can be independently performed to obtain for each input and output linguistic variable.
As for the corresponding standard deviations of the th membership function of , since and will be finely tuned in the reinforcement learning phase, we just use a first nearest-neighbor heuristic to estimate , which is given by
(26)
where
for
otherwise (27)
and is called an overlap parameter used to describe the degree of overlapping with two membership functions.
C. Reinforcement Learning Algorithm
Since there are no measurable output values fed back to in-struct the NFCRC to learn, a reinforcement learning algorithm is adopted and an evaluative handoff failure probability is used as a reinforcement signal. Fig. 3 also shows the diagram of the reinforcement learning for NFCRC, where the hierarchical cel-lular system provides the reinforcement signal as a desired output to NFCRC and receives the call admission control value and rate control value from NFCRC. The reinforcement signal is here defined as
(28) where denotes the QoS requirement of the desired handoff failure probability and is the actually measured handoff failure probability at time .
The reinforcement learning is applied to adjust parameters of input and output membership functions optimally, according to the input training data, the reinforcement signal, and the fuzzy logic rules. It derives updating rules for the mean and the stan-dard deviation of the bell-shaped membership functions so as to minimize the error function, defined as
(29)
For each training data set, starting at the input nodes, the down-up operation can compute to obtain the actual outputs of call admission control and rate control , and consequently is measured. On the other hand, from the output node, the up-down operation is used to compute
for all hidden nodes, where is the adjustable parameters such as the mean and the standard deviation for the input and output bell-shaped membership functions. We adopt the general learning rule
(30)
where is the learning rate. In the following, we show the com-putations of layer by layer, starting at the output nodes, and use the bell-shaped membership functions with cen-ters and the width as the adjustable parameters for these computations.
Layer 5: The updating rule for can be obtained by
LO et al.: NEURAL FUZZY RESOURCE MANAGER FOR HIERARCHICAL CELLULAR SYSTEMS 1203
the updating rule for by
(32)
An error signal in this layer , propagated to the proceeding layer, is given by
(33)
Layer 4: In this layer, only the error signal needs to be computed. is derived as
(34)
Layer 3: As in layer 4, only the error signal needs to be computed as
(35)
Layer 2: The adaptive rule of is derived as
(36)
and the adaptive rule of becomes
(37)
where . if
is minimum in the th rule node’s inputs; , other-wise.
IV. SIMULATIONRESULTS ANDDISCUSSIONS In the simulations, a hierarchical cellular system with microcells constructed along the Manhattan streets is assumed, and the handoff behavior of users is characterized by a teletraffic flow matrix [3], defined as shown in the equation at the bottom of the page, where , , represents the probability of a handoff call originated in cell and directed to cell ,
, and denotes the probability of this handoff call directed
to the adjacent macrocell, . and
would be zero.
The number of mobile stations in each cell is assumed to
be 550, and , . Suppose
and for voice calls and and
for data calls. Low- and high-mobility users are generated in a ratio of 7:3, and the cell dwell time is exponentially distributed with mean 180 s (18 s) and 1440 s (144 s) for high- and low-mobility users in macrocell (microcells), respectively. The speed of mobile users is assumed to be uniformly distributed in the range of 0–40 km (40–80 km) for low- (high-) mobility users. We also assume that the mean unencumbered session duration is s for voice call and s for data call, and the patience time for queued voice (data) calls is in the range of 5–20 s. One hundred fifty channels are fixedly allocated to macrocell and
micro-cells with a pattern of .
If the OCA and CCA schemes are applied, the system re-serves a number of channels as guard channels for handoff calls in cell , , which are denoted by . We do some simulations and obtain the
appropriate for OCA
scheme and for CCA
schemes at . Since the reneging (dropping) process is considered, it is not necessary to provide a large buffer size for new and handoff calls [9], [20], [21]; all buffer sizes in macrocell and microcells are assumed to be three. Note that in the following performance comparison, the OCA scheme provides no buffer and the CCA and FCAC schemes support the same buffering scheme and capacity as NFRM does.
Based upon the QoS requirement and knowledge of CCA [20], [21] and FCAC [9] mechanisms, parameters of membership functions for input linguistic variables
in the FCS are selected as follows: ,
, , and
for , , , ,
, and in (9)–(14); ,
, ,
and , for , , and
in (3)–(5); , ,
, and , for
, , and in (6)–(8).
In NFCAP, the initial values of membership functions of term sets for are chosen according to QoS requirement and then properly adjusted via the learning algorithm. Thus, the mean value in the membership function of
of is set to be 0.05 (0.02, 0), and let . In order to utilize the resource as much as possible and to guar-antee the QoS requirement, the initial values of membership functions of , , and for are set to be ,
, and ( , , and )
if the call is assigned to use the channels in macrocell
(micro-cell), and let and
.
The initial membership functions of the mean of the term set are set to be equally spaced in the range of [0,1], and let . The decision threshold in (23) is set to be for handoff call and for new call because handoffs are given higher priority than new calls. The use of dif-ferent may drastically reduce the training time required in the learning phase. As for and , their initial membership functions were heuristically set and required further optimiza-tion in the learning phase. Thus, was used.
Five performance measures such as the system utilization, the new-call blocking probability, the handoff failure probability, the forced termination probability, and the handoff rate are ob-served. The system utilization at time , denoted by , is de-fined as
(38)
where is the average number of busy channels in macrocell 0 (microcell ) at time and is the channel capacity for macrocell 0 (microcell ). The new-call blocking probability at time , denoted by , is defined as
(39)
where is the number of blocked (reneging) new calls in cell and is the number of new calls originating in microcell (macrocell-only region), at time
Fig. 4. U(t) for NFRM, FCAC, OCA, and CCA schemes.
.Similarly, the handoff failure probability at time , denoted by , is given by
(40)
A call will be forced termination if it is corrupted due to a handoff failure during its conversation time. The forced termi-nation probability at time , denoted by , is defined as
(41)
where is the number of blocked (dropped) handoff calls in cell and is the number of admitted new calls originated in cell , at time . The handoff rate at time
, denoted by , is defined as
(42)
Fig. 4 shows the system utilization versus the calling rate per user for schemes of NFRM, FCAC, OCA, and CCA at time . It reveals that the system utilization of NFRM gains 31.1%, 6.3%, and 1.4% improvement over the OCA, CCA, and FCAC methods, respectively. The superior performance of NFRM is because FCS in NFRM refers much more effective information than other conventional controllers, and it adopts fuzzy logic theory to balance traffic load between macrocell and microcells and provide a soft and accurate con-trol during traffic fluctuation. In addition, NFCRC in NFRM possesses the learning capability of neural networks to reduce the decision error and the fuzzy logic theory to qualitatively represent control rules naturally in the neural network to overcome some uncertainty and imprecision, and NFCRC contains the rate control function, which has the flexibility of rate assignment.
Fig. 5 shows the new-call blocking probability and the handoff failure probability for schemes of NFRM,
LO et al.: NEURAL FUZZY RESOURCE MANAGER FOR HIERARCHICAL CELLULAR SYSTEMS 1205
Fig. 5. P (t) and P (t) for NFRM, FCAC, OCA, and CCA schemes.
Fig. 6. P (t) for NFRM, FCAC, OCA, and CCA schemes.
FCAC, OCA, and CCA versus the calling rate per user at time . It can be seen that, as varies, of NFRM and FCAC remains constant at around , denoting that the system QoS requirement is guaranteed; and of NFRM is minimal, denoting that the system utilization is maximum, com-pared to FCAC, OCA and CCA. This is because NFRM uses neural fuzzy control to the allocation of channels. Neural net-works have merits of ability to learn from examples and to cope with incomplete input data. Fuzzy logic is a soft logic that is appropriate to represent in determining if a given requirement constraint is complied or violated. This in effect removes the imposition of worse case assumption from the decision-making of channel selection. The neural networks used in fuzzy call ad-mission control and rate manager can effectively estimate the optimal call admission and appropriately allocate a number of channels for each call. The other schemes are inadaptive to de-termine the number of guard channels to maintain, but not to overprotect, the QoS requirement as the traffic load is fluctu-ating and the changing is unpredictable.
Fig. 6 shows the forced termination probability versus the calling rate per user for schemes of NFRM, FCAC, OCA,
Fig. 7. R (t) for NFRM, FCAC, OCA, and CCA schemes.
and CCA at time . It is found that of NFRM has a flat curve under 2%. The reason is that NFRM obtains the unchanged , shown in Fig. 5.
Fig. 7 shows the handoff rate versus the calling rate per user for schemes of NFRM, FCAC, OCA, and CCA at time . It reveals that NFRM has more handoff rate than OCA by an amount of 2%. The reason is that the design of NFRM is based on the knowledge of FCAC and CCA, which combines overflow, reversible, and underflow. Fortunately, the signaling overheads for these handoffs might not cost so much as those for conventional handoffs between macrocells since most of these handoffs occur in the same macrocell. It also reveals that NFRM achieves less handoff rate than CCA and FCAC by an amount of 14.9% and 6.8%, respectively. It is not only because of more information, such as the speed of mobile station considered in NFRM, but also because of the neural fuzzy logic control that can provide decision support and expert system with powerful reasoning and learning capabilities.
V. CONCLUDINGREMARKS
In this paper, we propose a neural fuzzy resource manager for hierarchical cellular systems providing multimedia services. The NFRM mainly contains a neural fuzzy channel allocation processor, which is designed to be a two-layer controller. There is a fuzzy cell selector in the first layer and a neural fuzzy call-admission and rate controller in the second layer. The FCS uses soft logic to determine which cell, macrocell or micro-cell, to serve a call with channels. Then the NFCRC adopts a five-layer neural network with fuzzy logic control to deter-mine whether the call is accepted or not and how many channels are allocated. Simulation results show that the proposed NFRM improves the overall channel utilization by an amount 31.1% higher than the OCA scheme, 6.3% better than the CCA scheme, and 1.4% larger than the FCAC scheme, while maintaining the QoS requirement. It still reduces the handoff rate by an amount of 14.9% under the CCA mechanism and 6.8% below the FCAC scheme, but increases the handoff rate by an amount of 2% over the OCA mechanism.
ACKNOWLEDGMENT
The authors thank the anonymous reviewers for their con-structive suggestion for improving the presentation of this paper.
REFERENCES
[1] E. Dahlman, P. Beming, J. Knutsson, F. Ovesiö, M. Persson, and C. Roobol, “WDCMA—The radio interface for future mobile multimedia communications,” IEEE Trans. Veh. Technol., vol. 47, pp. 1105–1117, Nov. 1998.
[2] A. Samukic, “UMTS universal mobile telecommunications system: development of standards for the third generation,” IEEE Trans. Veh. Technol., vol. 47, pp. 1099–1104, Nov. 1998.
[3] S. S. Rappaport and L. R. Hu, “Microcellular communication systems with hierarchical microcell overlays: traffic performance models and analysis,” Proc. IEEE, vol. 82, no. 9, pp. 1383–1397, Sept. 1994. [4] , “Personal communication systems using multiple hierarchical
cel-lular overlays,” IEEE J. Select. Areas Commun., vol. 13, pp. 406–415, May 1995.
[5] C. Chang, C. J. Chang, and K. R. Lo, “Analysis of a hierarchical cellular system with reneging and dropping for waiting new and handoff calls,” IEEE Trans. Veh. Technol., vol. 48, pp. 1080–1091, July 1999. [6] C. L. I, L. J. Greenstein, and R. D. Gitlin, “A microcell/macrocell
cel-lular architecture for low- and high-mobility wireless users,” IEEE J. Select. Areas Commun., vol. 11, pp. 885–891, Aug. 1993.
[7] E. Gelenbe, I. W. Habib, S. Palazzo, and C. Douligeris, Eds., IEEE J. Select. Areas Commun. (Special Issue on Intelligent Techniques in High Speed Networks), Feb. 2000, vol. 18.
[8] K. Shum and C. W. Sung, “Fuzzy layer selection method in hierarchical cellular systems,” IEEE Trans. Veh. Technol., vol. 48, Nov. 1999.. [9] K. R. Lo, C. J. Chang, C. Chang, and B. Shung, “A QoS-guaranteed
fuzzy channel allocation controller for hierarchical cellular systems,” IEEE Trans. Veh. Technol., vol. 49, no. 5, pp. 1588–1598, Sept. 2000. [10] P. R. Chang and B. C. Wang, “Adaptive fuzzy power control for CDMA
mobile radio systems,” IEEE Trans. Veh. Technol., vol. 45, pp. 225–236, May 1996.
[11] J. S. Kim, S. H. Park, W. Dowd, and N. M. Nasrabadi, “Cellular radio channel assignment using a modified Hopfield network,” IEEE Trans. Veh. Technol., vol. 46, pp. 957–967, Nov. 1997.
[12] K. Smith and M Palaniswami, “Static and dynamic channel assignment using neural,” IEEE J. Select. Areas Commun., vol. 15, pp. 238–249, Feb. 1997.
[13] P. H. Chan, M. Palaniswami, and D. Everitt, “Neural network-based dy-namic channel assignment for cellular mobile communication systems,” IEEE Trans. Veh. Technol., vol. 43, pp. 279–288, May 1994.
[14] N. Funabiki and Y. Takefuji, “A neural network parallel algorithm for channel assignment problems in cellular radio networks,” IEEE Trans. Veh. Technol., vol. 41, pp. 430–437, Nov. 1992.
[15] D. Kunz, “Channel assignment for cellular radio using neural networks,” IEEE Trans. Veh. Technol., vol. 40, pp. 188–193, Feb. 1991.
[16] Z. Ali, A. Ghafoor, and C. S. G. Lee, “Media synchronization in mul-timedia web using a neuro-fuzzy framework,” IEEE J. Select. Areas Commun., vol. 18, pp. 168–183, Feb. 2000.
[17] C. T. Lin and C. S. G. Lee, “Neural-network based fuzzy logic control and decision system,” IEEE Trans. Computers, vol. 40, pp. 1320–1336, Dec. 1991.
[18] , “Reinforcement structure/parameter learning for neural-network-based fuzzy logic control systems,” IEEE Trans. Fuzzy Syst., vol. 2, pp. 46–63, Feb. 1994.
[19] , Neural Fuzzy Systems. Englewood Cliffs, NJ: Prentice-Hall, 1996.
[20] K. R. Lo, C. J. Chang, C. Chang, and B. Shung, “A combined channel assignment strategy for hierarchical cellular systems,” in ICUPC 1997, pp. 128–132.
[21] , “A combined channel assignment strategy for hierarchical cel-lular systems,” Comput. Commun., vol. 21, no. 13, pp. 1143–1152, Sept. 1998.
Kuen-Rong Lo was born in Hsinchu, Taiwan,
R.O.C., in 1957. He received the B.S. degree in elec-tronics engineering from National Taiwan University of Science and Technology, Taipei, Taiwan, in 1984, the M.S. degree in computer science from National Tsing Hua University, Hsinchu, Taiwan, in 1989, and the Ph.D. degree in electronics engineering from National Chiao Tung University, Taiwan, in 2000.
From 1984 to 1996, he was with Telecommuni-cation Labs, Directorate General of Telecommunica-tion (DGT), Ministry of CommunicaTelecommunica-tions, Taiwan, as a Design Engineer,Task Leader, and then Project Leader. In July 1996, he joined the Telecommunication Laboratories Chunghwa Telecom Co., Ltd., Chung-Li, Taiwan, where he currently is in the Division of Wireless Communication Tech-nologies. He has been a Project Manager since 1999. His research interests lie in the broad areas of mobile communication systems and mobile value-added terminal equipments. Presently, his research is focused on electronic fleet man-agement systems and 3G terminal equipment technologies.
Chung-Ju Chang (S’81–M’85–SM’94) was born
in Taiwan, R.O.C., in August 1950. He received the B.E. and M.E. degrees in electronics engineering from National Chiao Tung University (NCTU), Hsinchu, Taiwan, in 1972 and 1976, respectively, and the Ph.D degree in electrical engineering from National Taiwan University (NTU), Taiwan, in 1985. From 1976 to 1988, he was with Telecom-munication Laboratories, Directorate General of Telecommunications, Ministry of Communications, Taiwan, as a Design Engineer, Supervisor, Project Manager, and then Division Director. There, he was involved in designing digital switching system, RAX trunk tester, ISDN user-network interface, and ISDN service and technology trials in Science-Based Industrial Park. In the meantime, he also acted as a Science and Technical Advisor for the Minister of the Ministry of Communications from 1987 to 1989. In 1988, he joined the Faculty of the Department of Communication Engineering and Center for Telecommunications Research, College of Electrical Engineering and Computer Science, National Chiao Tung University, as an Associate Professor. He has been a Professor since 1993. He was Director of the Institute of Communication Engineering from August 1993 to July 1995 and Chairman of Department of Communication Engineering from August 1999 to July 2001. Now, he is the Dean of the Research and Development Office in NCTU. He was an Advisor for the Ministry of Education to promote the education of communication science and technologies for colleges and universities in Taiwan since 1995. He is also acting as a Committee Member of the Telecom-munication Deliberate Body. His research interests include performance evaluation, wireless communication networks, and broadband networks.
Dr. Chang is a member of the Chinese Institute of Engineers (CIE).
C. Bernard Shung received the B.S. degree in
elec-trical engineering from National Taiwan University, Taiwan, R.O.C., in 1981 and the M.S. and Ph.D. de-grees in electrical engineering and computer science from the University of California, Berkeley, in 1985 and 1988, respectively.
He is currently a Senior Engineer Manager with Broadcom Corp., San Jose, CA. He was a Visiting Scientist with IBM Research Division, Almaden Re-search Center, San Jose, in 1988–1990, and later a Research Staff Member in 1998–1999. He was a Fac-ulty Member in the Department of Electronics Engineering, National Chiao Tung University. Hsinchu, Taiwan, in 1990–1997. In 1994–1995, he was a Staff Engineer with Qualcomm Inc., in San Diego, CA, while taking a leave from NCTU. His research interests include VLSI architectures and integrated cir-cuits design for communications, signal processing, and networking. He has published more than 50 technical papers in various research areas.
