Cost-Based Vertical Handover Decision Algorithm for WWAN/WLAN Integrated Networks

Next generation wireless communications are expected to rely on integrated networks consisting of multiple wireless technologies. Heterogeneous networks based on Wireless Local Area Networks (WLANs) and Wireless Wide Area Networks (WWANs) can combine their respective advantages on coverage and data rates, o ﬀ ering a high Quality of Service (QoS) to mobile users. In such environment, multi-interface terminals should seamlessly switch from one network to another in order to obtain improved performance or at least to maintain a continuous wireless connection. Therefore, network selection algorithm is important in providing better performance to the multi-interface terminals in the integrated networks. In this paper, we propose a cost-based vertical handover decision algorithm that triggers the Vertical Handover (VHO) based on a cost function for WWAN/WLAN integrated networks. For the cost function, we focus on developing an analytical model of the expected cost of WLAN for the mobile users that enter the double-coverage area while having a connection in the WWAN. Our simulation results show that the proposed scheme achieves better performance in terms of power consumption and throughput than typical approach where WLANs are always preferred whenever the WLAN access is available.


Introduction
Beyond 3G mobile networks are being defined as the integration of diverse access technologies [1,2]. One promising way to construct integrated wireless networks is to integrate IEEE 802.11 Wireless Local Area Networks (WLANs) [3] and Wireless Wide Area Networks (WWANs), such as Universal Mobile Telecommunications System (UMTS) by 3rd Generation Partnership Project (3GPP) [4] and Code Division Multiple Access 2000 (CDMA2000) by 3GPP2 [5]. These networks have characteristics that perfectly complement each other in terms of both their data rates and coverage. Additionally, they are both widely deployed already. As a result, WWAN/WLAN-integrated networks can provide their users with a wider range of services at a higher quality. In the integrated wireless networks; a Mobile Node (MN) can roam dynamically between different wireless networks. Thus, one of the most important issues for integrated wireless networks is the Vertical Handover (VHO), which refers to a handover between systems with different air interfaces.
VHOs are classified into two categories: a downward VHO (DVHO) is a handover to a wireless network with a smaller cell size and (usually) higher bandwidth per unit area; an upward VHO (UVHO) is a handover to a wireless network with a larger cell size and (usually) lower bandwidth per unit area [6,7].
The natural trend for VHOs has been to use smallcoverage, high-bandwidth networks, such as WLAN, whenever they are available, and to remain connected to the network for as long as possible. Hence, the switch to lowbandwidth networks, such as UMTS and CDMA2000, occurs only when a high-bandwidth network is not available [6]. This type of network selection scheme for WWAN/WLANintegrated networks is called the WLAN-first scheme, where WLANs are always preferred by all services whenever the WLAN access is available to possibly increase bandwidth and reduce costs compared to WLANs. The studies in [8,9] propose the WLAN-first scheme, in which the VHO decision is based mainly on Received Signal Strength (RSS).
The studies in [10][11][12][13][14][15][16] propose cost-based VHO schemes that consider several factors, such as user preferences and network conditions, in order to perform flexible VHOs. In the cost-based VHO schemes, MN selectively changes its connection from WWAN to WLAN based on the cost function, even though the WLAN access is available. The authors of [10] consider higher level parameters which fall in the transport and application layers in addition to the available access networks' characteristics. In [11], the handoff decision is made by the system, aiming to achieve a better performance in the blocking probability. The authors of [12] address the main issues that arise while implementing the interoperability mechanisms between two different radio access networks. Although these VHO decision algorithms have certain advantages, they do not consider power efficiency, which is one of the most important aspects of wireless networks. This is especially true of integrated wireless networks since MNs with more than one wireless interface consume more power than those with one wireless interface. The studies in [13][14][15][16] consider the power consumption as one of the metrics for deciding VHO. However, they do not provide detailed analytical models to estimate network conditions such as available bandwidth and average user throughput in addition to power consumption.
In this paper we propose a new vertical handover decision algorithm based on a cost function for WWAN/WLANintegrated networks. In the proposed VHO decision algorithm, when an MN moves from the WWAN-only area to an area that is covered by both WLAN and WWAN, DVHO is performed only if the cost of WLAN is less than that of WWAN. Our cost function considers data throughput, power consumption, and monetary cost. In addition, user preferences for these factors are considered to select an optimal network for each MN. To estimate how much throughput is obtained and how much power is drained to send and receive a bit in the target WLAN, we develop an analytical model based on Automatic Rate Fallback (ARF) [17], which is generally used for adaptive modulation schemes in IEEE 802.11 WLAN. Adaptive mod-ulation schemes have been studied extensively and advocated at the physical layer in order to enhance throughput by adaptively matching transmission parameters to timevarying wireless channel conditions [18,19]. Therefore, in order to benefit from the use of adaptive modulation, we assume that our VHO decision algorithm operates on IEEE 802.11 WLAN, which employs adaptive modulation.
The rest of this paper has the following structure. Section 2 explains the details of our proposed scheme, including an analysis of the expected available bandwidth and the expected power consumption for WLAN with the adaptive modulation technique. Section 3 shows the effectiveness of the proposed scheme using the results from thorough simulations of various scenarios. Section 4 presents our conclusions.

Overall Architecture for the Vertical Handover Decision
Algorithm. Figure 1 illustrates an overall heterogeneous networking system architecture for the proposed VHO decision algorithm. As shown in Figure 1, our system model is based on the generic architecture for integration of WWAN and WLAN defined in [20]. In this architecture, the WLAN is connected to the Serving General Packet Radio Service (GPRS) Support Node (SGSN) via the GPRS Interworking Function (GIF) which provides a standardized interface to the GPRS core network and virtually hides the WLAN peculiarities. The primary function of the GIF is to make the SGSN consider the WLAN as a typical GPRS access system. That is, GIF is in charge of interfacing the WLANs with the WWAN by connecting to all Access Points (APs) in the WLANs as well as to SGSN in the WWAN.  Figure 2: The flow chart of the proposed vertical handover decision algorithm.
can be maintained in the system shown in Figure 1 when the MN changes an access technology. Thus, when a VHO occurs, the packets destined to the MN can be rerouted at the SGSN by using the intra/inter-SGSN Routing Area Update (RAU) procedure defined in [20] without going through the reauthentication process. We add the network selection module in SGSN to implement the proposed VHO decision algorithm. In our integrating architecture, the network selection module in the SGSN not only collects all the information related to the vertical handover but also makes the decision about what network to which an MN connect-based on the results of a cost function. Therefore, GIF must provide the SGSN with information about its WLAN, which is required to calculate the cost for WLAN. Figure 2 shows the flow chart for the proposed vertical handover decision algorithm. As shown in Figure 2, when an MN enters the double-coverage area that is covered by both WLAN and WWAN, SGSN collects all the required information for calculating cost from both WWAN and WLAN. The collected information includes the current power consumption, data rate, and monetary cost of each bit for WWAN as well as the available modulation levels, Packet Error Rate (PER) at each modulation level, expected power consumption, and monetary cost of each bit for the target WLAN. Based on this information, SGSN calculates the cost for both networks using the proposed cost function and then decides whether the MN will perform DVHO or not. That is, if the cost for WLAN is less than that for WWAN, the MN will switch from WWAN to WLAN. The result of this decision is delivered to the MN, allowing the MN to connect to the best available network. On the other hand, UVHO is performed when the MN moves from the double-coverage area to the WWAN-only area based on RSS.
We now explain the construction of the cost function. The factors that are taken into account for our cost function are data throughput, power consumption, and monetary cost. The cost function for network I is defined as where f t (I), f p (I), and f m (I) denote the normalized variables for data rate, power consumption, and monetary cost in network I, respectively. In (1), μ t , μ p , and μ m denote the weights of each factor, which are set according to the user preference. The constraint pertaining to μ t , μ p , and μ m is given by μ t + μ p + μ m = 1. Note that network I improves with lower values of C(I). f t (I) is obtained as follows: where α t and β t are variables to adjust between f t (I), f p (I), and f m (I) of (1). Because the three factors have different ranges of values, each factor has adjusting variables. D(I) is the throughput when using network I. Because the MN uses WWAN as its current network, it knows the value of D(wwan) by measurement. However, the MN does not measure the value of D(wlan) directly, and so the value is obtained as follows: where B(wlan) is the expected available bandwidth of network wlan. If B(wlan) is smaller than the current throughput, D(wwan), then the expected throughput, D(wlan), should 4 EURASIP Journal on Wireless Communications and Networking be B(wlan) because the network cannot send or receive more than B(wlan). If the available bandwidth is larger than current throughput, however, then the MN can send or receive more than D(wwan). This means that the value of D(wlan) is between B(wlan) and D(wwan). To apply this feature of the heterogeneous network to applications with different bandwidth requests, we introduce ρ, the value that represents dependency between the throughput requested by application and the network bandwidth. ρ is given between 0 and 1. It is dependent on current application that is using the network. If the MN uses real-time (rt) UDP applications, then the value is close to 0. If it uses greedy Best-Effort (BE) TCP applications, then the value is close to 1. For example, rt UDP applications such as VoIP or video call do not need to increase network throughput, although they are connected to a high-bandwidth network. That is, such applications occupy certain bandwidth as addressed in [21]. Thus, ρ is close to 0. If the MN has been using an rt UDP application in the WWAN-only area, that means that the WWAN admitted the MN's rt traffic by considering the WWAN's load status and checking whether the WWAN can allocate a certain bandwidth requested by the application (i.e., D(wwan)). If BE TCP applications like FTP are used, however, then the throughput, that can be achieved by the applications, is dependent on the available bandwidth [21], and so ρ is close to 1. On the other hand, if the MN' BE TCP traffic was admitted in the WWAN with high load, D(wwan) must be small for the MN, and hence, the throughput for the MN's BE application may be increased by handing over to the WLAN. The value of B(wlan) is calculated in Section 2.2. We now calculate the values of f p (I) and f m (I). Let pw(I) and m(I) be the power consumption and monetary cost for a user in network I, respectively. The expected amount of power consumption for the target WLAN, pw(wlan), is derived analytically in a later section, and m(I) can be obtained by Price(I)/D(I), where Price(I) indicates the price to transceive data in network I per unit time. We can then determine the normalized amount of power consumption as follows: The normalized monetary cost is defined as α p , β p , α m , and β m in (4) and (5) represent the adaptation variables. Note that we take the inverse number for m(I) and pw(I) in (4) and (5), respectively, since lower values are better. Whenever an MN enters the double-coverage area of WWAN and WLAN in the proposed vertical handover decision algorithm, SGSN analyzes the user preference so that μ t , μ p , and μ m in (1) are set. Based on these weights as well as the information collected from WLAN and WWAN, then SGSN calculates the costs for both WLAN and WWAN, C(wlan) and C(wwan), respectively. The MN is allowed to perform DVHO if the estimated cost for WLAN is less than that for WWAN. Otherwise, the MN maintains a connection with the WWAN.

Expectation for Available Bandwidth and Power Consumption.
In this section, we develop an analytical model to estimate the amount of power that is drained to send or receive a bit in the target WLAN. We start by establishing an analytical model of the available bandwidth for WLAN assuming that n-level modulation is allowed. We then estimate the amount of power consumption using this analytical model.

Available Bandwidth Expectation for WLAN.
In order to estimate the available bandwidth for WLAN, we determine the probability that the modulation is escalated after the transmission of k packets. This probability can be expressed as a function of the probability that a packet transmission succeeds, q, and the number of transmitted packets, k. Let f u (k, q) be this probability. Similarly, let f d (k, p) denote the probability that the modulation is lowered after the transmission of packet k, where p is the probability that a packet transmission fails, that is, packet error rate (PER). For analysis, we assume that the wireless channel is a Gaussian channel, in which each bit has the same bit error probability, and bit errors are identically and independently distributed over the whole frame as in [22].
Let N be the number of successfully transmitted packets that is required to elevate the modulation level, and let K be the number of unsuccessfully transmitted packets that is required to degrade the modulation level. That is, the next higher modulation level is selected after transmitting consecutive N frames successfully while the next lower modulation level is selected after failing to transmit consecutive K frames. Depending on the value of k, f u (k, q) can be decomposed into four cases as follows.
(1) When k < N. The modulation remains unchanged since consecutive N packets are not successfully transmitted, resulting in f u (k, q) = 0.
(2) When k = N. All N packets in the transmission should succeed so that f u (k, q) is computed as q N .
(3) When k = N + 1. In this case, the first packet transmission should fail, followed by N consecutive successful packets, resulting in pq N .
(4) When k > N + 1. As shown in Figure 3, we can divide this case into two components. In the first component, the modulation should remain unchanged during the transmission of the first k − N − 1 packets. Let A u (k − N − 1, q) be this probability. In the second component, the (k−N)th packet transmission should fail, followed by the successful transmission of N consecutive packets, resulting in pq N . Thus, 1st component 2nd component Figure 3: The case that the modulation is elevated after the transmission of k packets when k > N + 1.
A failed packet before successive N packets Move to lower modulation Move to higher modulation Successful packet Failed packet In order to obtain A u (k − N − 1, q) in (6), we subtract the complement probabilities from 1 instead of calculating them directly. The complement events, which mean that the modulation is changed during the transmission of the first k−N −1 packets, consist of the three cases shown in Figure 4. Case 3. It is also possible that the modulation remains unchanged after transmitting k − N − 1 packets, but after sending the K − Nth packet, the modulation degrades. In this case, the failure of K sequential packets occurs after the first phase, as shown in Figure 4. It is not included in Cases 1 and 2. f d (k − N, p) is the probability that modulation falls at k −N th packet. However, the last packet, or k −N th packet, is not included in the first phase, and so the probability of this case is ( By summing up all of the above three cases, we have Using the opposite derivation process, we can also obtain f d (k, p) as follows: where Using (6) and (7), we can calculate the average number of packets until the modulation is altered when the current modulation is given. Let g u (q) and g d (p) denote the average numbers of packets until the modulation is escalated and lowered, respectively. Along with f u (k, q) and f d (k, p), we have A key observation found in the ARF algorithm is that the transmission rate is always switched to adjacent one, so that the rate adaptation procedure of ARF could be expressed via a birth-death Markov chain as in [23]. Figure 5 shows the Markov chain where the state i represents the modulation level i of the single target station. Then, we can model the nlevel modulation technique as a simple queueing system with the state of the modulation level as shown in Figure 5, where λ i and μ i denote the rates that the modulation is elevated from M i to M i+1 and lowered from M i+1 to M i , respectively. λ i and μ i can then be expressed as functions of g u (q i ) and g d (p i ), respectively: where q i and p i represent the success and failure probabilities, respectively, when the modulation M i is used. From (9) and (10), the steady-state probability that the modulation level i is used, f (i), can be derived as where n i=1 f (i) = 1. We assume that B i and p i are the maximum bandwidths provided by the target WLAN and the PER at the modulation level i, respectively. Finally, we obtain B wlan , the expected bandwidth of the WLAN, as follows: Let θ wlan be the current usage of the WLAN. θ wlan is defined as the fraction of time during which the AP finds that the channel is busy for unit time: it can be obtained by the AP of the WLAN as in [21]. Then, we obtain the expected available bandwidth of the WLAN, B(wlan), as follows:

Estimation of Power Consumption.
In this section we estimate the amount of power consumption for both WLAN and WWAN. Let D in and D out denote the receiving and sending bit rates in WWAN, respectively. We then obtain the average amount of power consumption required to send or receive a single bit in WWAN, pw(wwan), by measuring the average amount of power consumption per unit time, δ wwan , as follows: As done in WWAN, we denote the receiving and sending bit rates in WLAN by D in and D out . Let δ r wlan , δ t wlan , and δ i wlan denote the amount of power consumed per unit time to receive, send, and stay-in-idle mode in WLAN, respectively. We can then obtain the average amount of power consumption in WLAN as follows: where B wlan is the available effective bandwidth for WLAN and can be obtained as derived in Section 2.2.1.
Here we assume that the bit rate in WLAN is affected by the difference in overall network bandwidth between WLAN and WWAN as well as the characteristic of ongoing applications. The amount of data throughput in WLAN, D , is then computed as where ρ represents the implementation variable based on the characteristic of application, as mentioned earlier (0 ≤ ρ ≤ 1). Assuming that D in and D out are adapted to the increased/decreased bandwidth with the same ratio, our proposed scheme estimates a reduction in power consumption by using (14), as well as (15), if the MN performs DVHO.

Performance Evaluation via Simulation
We have developed the simulator using C++ to investigate the effectiveness of our proposed scheme in comparison with that of the existing WLAN-First Scheme (WFS) and a cost-based scheme [14] where a cost function of power consumption, available bandwidth, and monetary cost, is used for network selection. The proposed scheme takes account of the MN's RSS and ARF when computing the cost for each network, whereas the cost-based scheme does not. Our simulation model accounts for the throughput and power consumption as performance metrics. The number of VHOs is also evaluated.

Simulation Environment.
In our simulation, we use the network topology illustrated in Figure 1 in which five APs are deployed in a single WWAN coverage of 500 × 500 m 2 . As shown in Figure 1, each AP in the network topology covers a circular area with a radius of 100 m and employs IEEE 802.11a standard protocol [24], including an adaptive modulation technique. In our simulation, the bandwidth for WLAN is set to a maximum of 54 Mbps, while the WWAN uses the bandwidth of High-Speed Downlink Packet Access (HSDPA), 14 Mbps. The data rate for WLAN changes adaptively based on an adaptive modulation technique through which five modulation levels are available. Table 1 shows the characteristics of the five modulation levels for 802.11a adopted in our simulation [19,24]. Both K and N for adaptive modulation are set to 10. In our simulation, the PER at each modulation level i for WLAN is modeled in [19,Equation (5)] as follows: where γ represents the received signal-to-noise ratio (SNR). The level-dependent parameters a i , g i , and γ pi are obtained by fitting (17) to the exact PER. Supposing that a packet length is 1080 bits, then Table 1 provides the fitting parameters for transmission modes. Assuming that the path-loss exponent is 3, the received SNR, γ, is computed as follows [18]: where P t is the transmit power, PL(d0) = 46.7 dB is the mean path loss at a close reference distance of d0 = 1 m, d is the distance in meters, and G(t) is a time-variant Rayleighdistributed fading gain, for which the well-known Jakes' Doppler spectrum is assumed. The Correspondent Node (CN) generates two different classes of traffic: BE TCP traffic (ρ = 1) and UDP traffic (ρ = 0). The data session arrival rate follows a Poisson process with mean 1/300. The packet size is 1000 bytes. In the simulation, TCP Reno has been implemented, where the initial and maximum window sizes are set to 1 and 1024 packets, respectively. Each session generates a file with the size varying from 1 to 11 Mbytes. To simulate the UDP traffic, On/Off traffic is generated, where the On and Off periods are exponentially distributed with means of 100 seconds and 300 seconds, respectively. During On period, traffic is generated with a rate of 20 to 200 Kbytes/s. We obtain the values of (α t , β t ), (α p , β p ), and (α m , β m ) in (2), (4), and (5) from the simulation such that they normalize the value of each factor, that is, f t (I), f p (I), and f m (I). Table 2 gives these values. In our simulation, the  For the mobility of MNs, we use the Smooth Random Mobility model, in which there is a maximum speed, v max m/s, and a set of two preferred speeds, v pre f1 = (3/5)v max and v pre f2 = v max [25]. In our simulation, a total of six MNs move at pedestrian speeds. The initial velocity of each MN is chosen from the two preferred speeds, v pre f1 , v pre f2 , and the range [0, v max ] with probabilities of 0.2, 0.5, and 0.3, respectively. We assume a uniform distribution on the range of [0, v max ]. All the MNs move with the selected speed until a new target speed is set again. The target speed is updated at every time interval which follows an exponential distribution with a mean of T v s. Whenever a new target speed is decided, the MN decelerates or accelerates until the target speed is reached or a new target speed is decided again. a min and a max denote the maximum possible deceleration and acceleration in m/s 2 , respectively. The MN selects a random acceleration/deceleration value from [a min , a max ]. It also decides whether to change its movement direction or to maintain it with probabilities of 1/6 and 5/6, respectively. This is done at every time interval with an exponential distribution, the mean of which is T ϕnew s.
All the parameter values for the mobility model are obtained from [25,26] and are given in Table 3. In the simulation, MNs move in accordance with two input sets that represent two different levels of mobility, as shown in Table 3. That is, half of the MNs move according to Input 1, and the other half of the MNs move according to Input 2.

Simulation Results.
We first run simulation tests by setting the weights (μ m , μ t , μ p ) in (1), which represent user preference, all to 1/3 for the proposed scheme. To determine the impact of the weights in (1) on the performance of the proposed scheme in terms of throughput, power consumption, and monetary cost, we repeat the simulation with Cost-based scheme WFS Figure 6: File size per session versus throughput when the TCP traffic (ρ = 1) is generated. different combinations of μ m , μ t , and μ p . We run simulation tests for the proposed scheme with another three weight sets that make the data throughput, the power consumption, and the monetary cost most important, respectively, among the three factors in our cost function. The three weight sets are The values for performance metrics in simulation results are those averaged over all the MNs. Figure 6 shows the throughput performance for different file sizes when the TCP traffic (ρ = 1) is generated. We observe from Figure 6 that both the proposed and the costbased schemes achieve much better throughput performance than WFS over the whole range of the file size. Specifically, the throughput improvement by our scheme over WFS is 17.3% in average. Table 4(a) shows the relative throughput performance gains of our scheme with the four weight sets and the cost-based scheme over WFS when ρ is 1. We see that when a higher weight is put on the throughput (i.e., μ t = 0.8), our scheme shows the highest throughput improvement. It can be also seen from Table 4(a) that the throughput is also improved by the proposed scheme over the cost-based scheme. Figure 7 shows the power consumption per minute in Watts versus the different file sizes when the TCP traffic (ρ = 1) is generated. We see from Figure 7 that both the proposed and the cost-based schemes show less power consumption than WFS for all the values of the file size. It is also observed from Figure 7 that our scheme consumes less power than the cost-based scheme except for the case of μ m = 0.8, in which the power consumption for the proposed scheme is almost same as that for the cost-based scheme since 80% of the weight is put on the monetary cost in the proposed scheme. Specifically, Table 4(b) shows the relative performance gains of our scheme with the four weight sets and the cost based scheme over WFS in terms of the power consumption when ρ is 1. As expected, the proposed scheme with μ p = 0.8 shows the most performance gain. Our scheme decreases the power consumption per minute from those of WFS and the cost-based scheme by 76.71% and 35.35%, respectively, in average. Figure 8 shows the number of VHOs per hour for the different file sizes when the TCP traffic (ρ = 1) is generated. We see that the WFS yields a larger number of VHOs, compared to both the proposed and the cost-based schemes. In the proposed scheme, the average number of VHOs performed by each MN is 85.51% less than that of WFS in average. This is because WFS lets each MN perform DVHO whenever the MN enters the double-coverage area. The costbased scheme shows almost the same number of VHOs as the proposed scheme with μ m = 0.8, whereas our scheme performs 73.7% less VHOs than the cost-based scheme, in average.
From Figures 6, 7, and 8, we can see that when the TCP traffic is generated, our scheme outperforms WFS in terms of the throughput and the power consumption for the different file sizes, while yielding a smaller number of VHOs than WFS. It can be also seen that the proposed scheme achieves better performance than the cost-based scheme in terms of the throughput, the power consumption, and the number of VHOs, except for the case in which 80% of weight is put on the monetary cost.
We now evaluate the performance of our VHO decision algorithm when the UDP traffic (ρ = 0) is generated. Table 5 shows the throughput performance for different  values of the data transmission rate when ρ = 0. We run simulation tests for the proposed scheme with the four weight sets, as in the simulation for the TCP. We observe from Table 5 that the proposed and the cost-based schemes and WFS have similar throughput performances, because when ρ is 0, the amount of data throughput achieved by WWAN is same as that achieved by WLAN, as known from (16). UDP traffic (ρ = 0) is generated. We see from Figure 9 that our scheme with μ p = 0.8 achieves the greatest improvement over the WFS and the cost-based scheme in terms of the power consumption. Specifically, our scheme with μ p = 0.8 decreases the power consumption per minute 83.71% and 68.35% from those of WFS and the cost-based scheme, respectively, in average. For the other three weigh sets (i.e., μ t = 0.8, μ m = 0.8, and all the weights are 1/3), our scheme consumes 1.33% less power than the cost-based scheme in average, because less weight is put on the power cost, compared to the case of μ p = 0.8. Figure 10 shows the number of VHOs per hour for the different values of the data transmission rate when the UDP traffic (ρ = 0) is generated. We see that WFS, in which the MN performs DVHO whenever it enters the double-coverage area, shows the greatest number of VHOs. The number of VHOs for the cost-based scheme is 2.8% more than our scheme with the three weight sets (excluding μ p = 0.8), in average. However, for our scheme with μ p = 0.8, no VHO is performed over the whole range of the data transmission rate in order to comply with the user's request and to save power as much as possible.

Conclusion
The coexistence of different wireless access technologies including WLAN and WWAN (e.g, UMTS and CDMA2000) has enabled for multi-interface MNs to roam freely from one network to another, being able to maintain their network connection and the QoS required by higher layer applications. In the integrated WWAN/WLAN network, the MNs should hand over from one network to another in order to obtain improved performance or at least to maintain a continuous wireless connection. Therefore, in this paper we proposed a cost-based vertical handover decision algorithm that operates based on a cost function for WWAN/WLAN integrated networks. In the proposed scheme, when an MN moves to the double coverage area, DVHO is performed only if the cost expected in the WLAN is less than that in the WWAN. The cost function is based on data throughput, power consumption, and monetary cost. In addition, user preferences for these cost factors are considered to select an optimal network for each MN. Our simulation results showed that the proposed scheme outperforms the typical WFS scheme in terms of throughput and power consumption by selecting a lower cost network over all user preferences.