{christina.werner;jens.voigt}@actix.com. ABSTRACT. The increasing traffic and the demand for high data rate ser- vices in WCDMA networks prompt the need for ...
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07)
HANDOVER PARAMETER OPTIMIZATION IN WCDMA USING FUZZY CONTROLLING Christina Werner*, Jens Voigt*, Shahid Khattak**, and Gerhard Fettweis** *Actix GmbH **Dresden University of Technology Altmarkt 10, D-01067 Dresden, Germany D-01062 Dresden, Germany {christina.werner;jens.voigt}@actix.com A BSTRACT The increasing traffic and the demand for high data rate services in WCDMA networks prompt the need for an automatic network optimization. Current state–of–the–art optimization methods adapt physical parameters, such as the antennas’ tilt or azimuth. As agreed in the community, a further capacity increase can be achieved by an optimization of Radio Resource Management (RRM) parameters. While theoretical approaches of RRM parameter optimization have already been introduced in the literature, practical solutions optimizing cell individual parameters have rarely been treated so far. Consequently, this paper copes with the optimization of the cell individual offset (CIO) in the soft handover algorithm with the goal to reduce the network’s outage probabiliy. We design and apply a fuzzy logic controller, whose output are CIO changes matching the current traffic and load conditions. Verifications using a dynamic system simulator prove the powerfulness of our method and promise significant outage reductions in a fully loaded network scenario leading to capacity improvements of up to 9%.
Keywords — WCDMA; Radio Resource Management; Soft Handover Control; Fuzzy Controlling; Simulation and Performance Evaluation; Automatic Parameter Optimization I.
I NTRODUCTION
The increasing traffic and the demand for high data rate services in WCDMA, as well as the network size and complexity, prompt the need for automatic optimization, whose aim is to save capital and operational expenditures. Current state–of– the–art optimization methods adapt physical parameters, such as the antennas’ tilt or azimuth and, possibly, the transmit power of the common channels [1]. As agreed in the community, a further capacity increase can be achieved by optimization of parameters of the radio resource management (RRM) algorithms, whose task is to dynamically allocate, maintain, and release the available hard (e.g. channelization codes) and soft (e.g. power) resources. As the default, the RRM parameters have fixed or even standard values given by the vendor, which can not match the current traffic, interference, and propagation conditions in a cell. In the start phase of WCDMA, an optimal setting of RRM parameters was not of a great importance, since the traffic in the third generation network was comparatively low. However, the traffic increase and the demand for high data rate services in a network with sub–optimal configurations currently leads to an increase in the blocking and dropping probabilities, whose sum is also called the outage probability. To prevent a reduction of c 1-4244-1144-0/07/$25.00 2007 IEEE
the user satisfaction and the operator revenue, and to increase the network capacity, a local optimization of the RRM parameters is inevitable and currently strongly demanded by network operators. Recent research on optimization of RRM parameters, such as admission and congestion thresholds, or soft handover reporting range and hysteresis has shown that an adjustment of the RRM parameters improves the system capacity in comparison to a network, whose RRM values are fixed, e.g. [2], [3], and [4]. Nevertheless, although already motivating cell individual parameter optimizations [4], these publications deal with RRM parameters, whose values can only be configured at the Radio Network Controller (RNC) and are not parameterizable at a cell individual level. Instead, we consequently optimize RRM parameters per cell depending on the temporary and local cell specifics. In order to show that an automated adjustment can improve the network capacity, the RRM parameter cell individual offset (CIO), which influences the soft handover (SHO) overhead in a cell, was chosen to be optimized in this paper. As the name already implicates, the value of this RRM parameter is cell–specific and the choice of its value can consider the cell characteristics. In particular, the CIO will be optimized with the goal to reduce the SHO overhead and further the outage probability in case of an overloaded network. As an optimization strategy, fuzzy logic controlling (FLC) is applied. FLC for the tuning of SHO parameters in CDMA networks (IS-95) was first introduced in [5]. A fuzzy logic concept for the dynamic tuning of admission thresholds and RNC– specific SHO parameters has been suggested in e.g. [3] and [4]. The authors showed that FLC is a suitable method for the RRM parameter tuning. This paper is organized as follows: Section II. provides the reasons for the choice of the CIO, which is the parameter the optimization in this work focuses on. In Section III. the fuzzy logic controller for the CIO is designed. In Section IV. the verification tool and setup are introduced. The results of our investigations are shown in Section V.. Finally, Section VI. gives our conclusions. II.
C ELL I NDIVIDUAL O FFSET C ONSIDERATION AND O PTIMIZATION TARGET
Since the Mobile Stations (MS) in WCDMA are separated by codes and share the same frequency, a MS has the possibility to be connected to more than one cell simultaneously. This state is termed soft handover (SHO). The cells to which a MS is connected during SHO is denoted as the active set.
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07)
SHO has the following advantages: • SHO ensures the seamless cell change and increases the network coverage on the uplink link direction (UL). • The SHO combining gain mitigates propagation effects, such as fast fading and shadowing. • The SHO combining gain might lead to a transmit power reduction, which reduces the interference in the network on the UL. On the other hand, SHO also has some disadvantages: • At each cell extra codes, hardware, and downlink (DL) power have to be allocated for the additional SHO links. This leads to a higher outage probability and a decrease in the network capacity. Especially the DL power is a very scarce resource, since it is shared by all MS in the cell. The trend for services with asymmetrical data rates and High–Speed Downlink Packet Access (HSDPA) intensifies the DL cell power deficiency. • SHO links cause additional DL interference for the remaining MS. The CIO is a cell–specific RRM parameter, whose value modifies the cell’s pilot measurements and thus, influences the time and the place of SHO establishment with the particular cell, see e.g. [6] or [7], §14.1. A change of the CIO value has the following influence on the network performance: Effects of a higher CIO: More MS add the cell to or keep it in their active set. Thus, a higher value of the CIO leads to a SHO overhead increase in the cell, stressing both, SHO advantages and disadvantages. An optimizer could preferably rise the CIO at rather low DL load values. Effects of a lower CIO: Less MS add the cell to their active set and at the same time more MS remove it from the active set. Thus, a lower value leads to a SHO overhead decrease, easing both, SHO advantages and disadvantages. An optimizer could preferably lower the CIO at high DL load values. Thus, carefully adjusted CIOs can help to adapt the SHO overhead to the current load situation in a cell. This process will result in lower outage probabilities in case of overload situations in the network, which is the optimization aim in this work. Fuzzy logic controlling (FLC) was chosen as a load– dependent optimization method for the CIO. The following section presents the designed and implemented fuzzy logic controller. III.
F UZZY C ONTROLLER FOR THE C ELL I NDIVIDUAL O FFSET
Fuzzy controlling is based on the principles of fuzzy logic, which maps, by the means of membership functions, linguistic terms, such as “very”, “high”, “low”, etc. not only to the values 0 and 1, but to the whole interval [0, 1]. FLC offers a simple software implementation. It is based on if–then rules, which express the relationship between the input and the output variables. The rules are derived from statements
that are formulated in natural language. FLC can be used for linear and non–linear control, but is especially suitable for non– analytical solvable non–linear problems. See e.g. [5] for a more comprehensive introduction of FLC. FLC design is based on expert knowledge and analysis. The lack of specific design criteria is the most important disadvantage of FLC. Therefore, a strong expert background and a thorough analysis of the system behavior is a prerequisite for the design of a good fuzzy controller. A.
FLC Architecture
A fuzzy logic controller contains a fuzzification module, an inference engine with a rule base, and a defuzzification module. The FLC structure is shown in Figure 1. Rule Base
Input
Inference Engine
Fuzzification
Defuzzification
Output
Figure 1: Structure of a Fuzzy Logic Controller (FLC) Based on the conducted analysis of the CIO modifications and on the FLC theory, a fuzzy logic controller for the CIO was designed. The construction steps include the choice of the input parameters, the linguistic regions as well as the corresponding membership functions, the rule base, the type of inference, and the defuzzification method and are described next. B.
Choice of the Input Parameters and Fuzzification
An FLC with two input parameters was considered in the first investigation stages to ensure low complexity and implementation effort. Among the large number of possible input parameters (i.e. DL load, SHO overhead, blocking and dropping rates, lack of coverage, mean active set size, ... ), the DL load and the SHO overhead were chosen because the aim of the controlling process is to reduce the SHO overhead in a cell if the measured DL load is high. 1)
DL Load
The DL load is measured as the ratio between the averaged total DL transmit power P t, total, DL and the cell’s maximum output power Pt, total, DL, max : MDL =
P t, total, DL Pt, total, DL, max
(1)
The averaged total DL transmit power includes the power of the common and dedicated channels. 2)
SHO Overhead
The second input parameter, the SHO overhead, is calculated as the ratio of the MS that contain the cell in their active set (this cell is not the best cell) (NDCH − NBest ) and the total count of MS on a dedicated channel in the cell (NDCH ): MSHO = 1 −
NBest NDCH
(2)
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07)
The fuzzification step converts the real input values into fuzzy sets. The membership degree of the measured input values to the considered properties is determined by means of the corresponding membership functions and is in the interval [0, 1]. In this paper triangular symmetric membership functions are used (compare [8]). The properties, which are also called linguistic regions, describe the measured values of the input parameters. One membership function is associated with one linguistic region. Both input parameters in the present work are described by three linguistic regions — “low”, “medium”, and “high”.
The output membership degrees for each linguistic region of the output parameter are then used as an input for the defuzzification step.
C.
In the present work, the network performance with optimized CIO values by the FLC is investigated first with the aim to identify further input parameters which should be considered by the FLC. Therefore, single services are investigated first. Based on the obtained results, the FLC will be extended and applied then in various multiple–services scenarios. Finally, the FLC will be used for optimization in a real network scenario. The investigation tool and the obtained results are discussed next.
Inference Engine and Rule Base
In the second step, the inference engine applies the rules in the rule base and outputs a fuzzy value. The rules in the rule base are designed using the experience and knowledge of the constructor of the FLC. Their number depends on the number of the input variables and on the number of the distinguished linguistic regions per input variable. 1)
Rule Base
Due to the fact that three linguistic regions per input parameter have been defined and two input parameters are used, the resulting rule base is two–dimensional and contains nine rules. The rule base is shown in the form of a matrix in Figure 2. It is based on ideas presented in [8]. MDL
Low
Medium
High
Low
High
Medium
Low
Medium
High
Medium
Low
High
Medium
Low
Low
MSHO
Figure 2: Rule Base for the CIO FLC It should be read as follows: The CIO is reduced at high DL loads or when a high portion of MS in the cell is in SHO in order to release resources and to reduce the outage probabilities. When the DL load is low and the portion of MS in SHO is not high, then the CIO can be increased to use the combining gain and improve the quality and coverage. 2)
D.
Defuzzification
In the following defuzzification step, the fuzzy output values are converted into real numbers. The defuzzification method used in this work is the Weight–of–Average–Formula [8]. The advantage of this defuzzification method in comparison to other frequently used defuzzification methods is that it considers all output linguistic regions.
IV.
E XPERIMENTAL S ETUP
The investigations were carried out using a dynamic system– level simulator, which allows a detailed time–dependent analysis of the network performance considering multi–user and multi–service scenarios [9]. In the first investigation stage, an ideal hexagonal cell environment area was used. The simulation area includes 19 Node Bs (= b 57 cells) and its total area is equal to 21 km2 . In order to exclude border effects from the simulation results, we only evaluated the results of the inner circle of cells. Its area is 7 km2 and it contains 7 Node Bs (= b 21 cells). The distance between the Node Bs is equal to 1 km. The examined services possess a homogeneous traffic distribution. In this work only circuit–switched services are considered. The reason for this choice is that the Dedicated Channels (DCH), which are subject to SHO, are definitely required by the circuit–switched services. The main investigation parameters are shown in Table 1. Depending on the service and mobility profiles it can be distinguished between four MS profiles: SL (Speech, Low Speed), SH (Speech, High Speed), VL (Video, Low Speed), and VH (Video, High Speed). All performed investigations lasted 5 minutes real time (= b 30,000 UMTS frames).
Min–Max Inference
The FLC for the CIO uses min–max inference, which contains two steps — aggregation and composition. In the aggregation step, the minimum of the membership degrees of the two input values is determined for each rule. In the composition step, a membership degree for each linguistic region of the output parameter is calculated using the rule base and the values determined in the aggregation step. For the determination of the output membership degrees, the maximum function considers only the rules for which the linguistic region is contained in the rule output.
V. A.
R ESULTS AND D ISCUSSION
Single Services Analysis 1)
Performance Improvement
The FLC presented in Section III. was applied in two operating modes: Offline Mode: In the offline mode, the necessary input parameters for the FLC were extracted as average output results from a first simulation run applying standard values for the CIO in all cells. Then the FLC was used to determine average optimized
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07)
Table 1: Main Investigation Parameters Parameter
Value
Pt, total, DL, max Fast Fading Type Shadowing Std. Dev. Low Speed High Speed Speech Rate Video Rate SHO Add/Drop Window SHO Time to Trigger
43 dBm COST 207 Typical Urban 6 dB 1 m/s (average/normal distr.) 16 m/s (maximum/uniform distr.) 12.2 kbit/s 64 kbit/s 3/5 dB 640 ms
values for the CIO. These CIO settings were then applied in a second simulation run as constant values over the entire simulation run time. This operating mode emulates the real application scenario within network operators: Instead of getting input parameters out of a first simulation run, the network operator can use average networks counters e.g. for the busy hour to feed them into the FLC in order to get better CIO values for that specific network setup and load. Online Mode: In the online mode, the necessary input parameters for the FLC were extracted within a simulation run and given to the FLC. The output of the FLC, improved CIO values, were simultaneoulsy applied within the same simulation run. This operating mode emulates online optimization of a network, e.g. with the operating sub–system (OSS). The investigations started with a highly overloaded network setup having outage probabilities around 10% for the different service profiles. The outage probability is hereby measured as the relative number of MS, which could not be served due to missing coverage, blocking, or dropping. The outage probabilities for the three investigated cases — without FLC, in online mode, in offline mode — is shown in Figure 3. 25%
20.7
3.
8 4. 8% 74 %
SH
2.
SL
66 3. % 55 %
7.
3.
41 % 4. 69 %
9.
42 %
91 %
8% 11 .5 08 2. % 27 %
5%
2.
Outage
10%
2)
Adjustment of the FLC for the CIO
The evaluation of the single services investigations showed that different CIO values are optimal for different MS service profiles and speeds. The mean CIO values, which lead to a local minimum of outage probabilities for different service profiles, are presented in Table 2. Table 2: Average CIO values, which lead to the lowest outage MS Profile CIO (dB) SL −2.18 −1.10 SH VL −2.41 −1.05 VH Based on the obtained results, we suggest adding two additional binary input parameters to the FLC — the average speed and data rate of the MS in a cell. The improved FLC (with two further input parameters) was then applied in a situation with multiple services. B.
Multiple Services Analysis
The improved FLC was used to optimize the CIO value in three different multiple services scenarios: 1. prevailing MS with low velocity and low data rate, 2. prevailing MS with high velocity and low data rate, and 3. high portion of MS with low velocity and high data rate. The share of each MS profile is shown in Table 3. The results lead to the conclusion that it is advisable to set the CIO value for the prevailing type of service in the cell to achieve the lowest outage.
1%
20% 15%
yielded an average outage decrease of 63 %. Thus, the replacement of the time–dependent CIO calculated by the FLC with its average leads to a slight increase of the outage. In the following sections, only offline optimization is considered. The outage improvement is greater for the service profiles SL and VL. The reason for this is the low velocity of these profiles, which allows for a further CIO reduction.
Table 3: Multiple services composition SL SH VL M-SL 60 % 20 % 20 % M-SH 20 % 60 % 20 % – 40 % M-VL 60 %
0% VL
VH
UE Profile without Fuzzy
Fuzzy-Online
Fuzzy-Offline
Figure 3: Outage Reduction with Online and Offline FLC Optimization By applying the FLC for the CIO, the outage probability could for all service types be reduced to under 5 %, which is an acceptable value from the view of most network operators. The achieved outage decrease (online) is between 54 % (SH) and 87 % (VL), on average — 71 %. FLC offline optimization
The network capacity, in terms of traffic, where the outage probability was almost the same with and without the FLC, was now examined. The capacity gain is shown in Figure 4. The application of the FLC leads to an increase of the MS count, which can be added to the network without an increase of the outage probability of less than 5%. For the multiple services scenario MVL, the relative capacity gain is lower because the user count, irrespective of the MS profile type and data rate, is considered here. On average over all MS profiles, a user gain of about 9% could be observed.
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07) 116.02%
120%
108.48%
104.86%
User Gain
100% 80% 60%
• the DL cell load,
40%
• the SHO overhead,
20%
• the data rate of the dominant user type in the cell, and
0% M-SL
M-SH without Fuzzy
M-VL Fuzzy-Offline
Figure 4: User Gain at an Outage Probability less than 5 % with FLC Offline Optimization (multiple services) C.
The design of the FLC requires a thorough prior analysis of the parameters’ impact on the network performance. In particular, we optimized the RRM parameter CIO of the SHO algorithm with an FLC. As input to the controller we suggest to use the following averaged busy hour network counter directly or as derived values:
Real Network
Finally we investigated the powerfulness of our algorithm in a real network environment that was a part of a major European city center. The simulation area covered 69 km2 and the analysis area 15 km2 . The simulation area contained 43 Node Bs (= b 129 cells) and the analysis area included 14 Node Bs (= b 42 cells). The available clutter, elevation, and traffic data were used to ensure a precise modeling of the environment and propagation conditions in the real network. Furthermore, the antennas’ azimuth and tilt had already been optimized. The multiple services composition contained mostly the SL service profile. The investigations were again started with an overloaded situation. The performed offline CIO optimization by the FLC reduced the outage from 7.5 % to 4.8 %, which corresponds to an improvement of 36 %. The lower CIO values did not sacrifice the connection quality. Not only the target block error rate of 1 % was maintained, but even the Ec /I0 was inreased by 3 %. The number of control information units such as MS measurement reporting events 1A, 1B, and 1C (see [7]) could also be reduced due to the optimized CIO values, further reducing the network load. To summarize, the investigations of the real network proved that the optimization with the FLC, as described in this work, improves the network performance and reduces the outage probability. VI.
C ONCLUSIONS
Our results imply that in the case of an overload situation the default CIO value (0 dB) leads to an unnecessary SHO overhead, which consumes DL power that could be granted to new users’ requests. Thus, the user satisfaction and the operator revenue can be increased by the choice of cell–individual values of selected RRM parameters. A suitable automated optimization solution is the application of an FLC, which optimizes the RRM parameter values depending on the cell’s characteristics.
• the speed of the dominant user type in the cell. In addition to a tremendous reduction of the outage probability in a fully loaded network scenario, the application of the developed FLC promises to lead on average to up to a 9 % increase of the real multiple services network capacity in comparison to a network with standard CIO values. A simultaneous online optimization of the CIO within the operating network could even outperform the investigated offline busy–hour optimization. ACKNOWLEDGEMENT We would like to thank our colleagues at Actix GmbH and at the Vodafone Chair for Mobile Communications Systems at Dresden University of Technology for their support and reviews. R EFERENCES [1] M. J. Nawrocki, M. Dohler, and A. H. Aghvami, Understanding UMTS. Radio Network Modelling, Planning and Automated Optimisation. John Wiley & Sons, 2006. [2] A. Hoeglund and K. Valkealahti, “Automated optimization of key WCDMA parameters,” Wireless Communications and Mobile Computing, vol. 5, no. 3, pp. 257–271, 2005. [3] J. Picard, H. Dubreil, F. Garabedian, and Z. Altman, “Dynamic control of UMTS networks by load target tuning,” in Proc. IEEE Vehicular Technology Conference (VTC-Spring), vol. IV, (Milan, Italy), pp. 2351–2354, May 2004. [4] R. Nasri, Z. Altman, H. Dubreil, and Z. Nouir, “WCDMA downlink load sharing with dynamic control of soft handover parameters,” in Proc. IEEE Vehicular Technology Conference (VTC-Spring), vol. II, (Melbourne, Australia), pp. 942–946, 2006. [5] B. Homnan, V. Kunsriruksakul, and W. Benjapolakul, “Adaptation of CDMA soft handoff thresholds using fuzzy inference system,” in Proc. IEEE International Conference on Personal Wireless Communications (PWC), (Hyderabad, India), pp. 259–263, Dec. 2000. [6] H. Holma and A. Toskala, WCDMA for UMTS. John Wiley & Sons, 3rd ed., 2004. [7] “Radio Resource Control (RRC) – Protocol Specification,” in 3GPP TS25.331, version 6.12.0, (Sophia Antipolis Cedex, France), ETSI, 2006. [8] B. Homnan and W. Benjapolakul, “QoS controlling soft handoff based on simple step control and a fuzzy inference system with the gradient descent method,” IEEE Transactions on Vehicular Technology, vol. 53, pp. 820– 834, May 2004. [9] W. Rave, T. Koehler, J. Voigt, G. Fettweis, P. Schneider, and M. Berg, “Evaluation of load control settings in an UTRA/FDD network,” in Proc. IEEE Vehicular Technology Conference (VTC-Spring), vol. IV, (Rhodes Island, Greece), pp. 2710–2714, May 2001.