Abstract
Network selection is a common issue faced by mobile users. There are several approaches that can be used to optimize the quality of service (QoS) in wireless local area network (WLAN) and cellular network namely by connecting to optimal access point (AP) or using parallel path at transport layer such as transmission control protocol (TCP). In this paper, progressive mobility prediction (PMP) is used to predict the optimal WiFi AP that mobile devices should be connected to for better connectivity. In PMP, dual hidden Markov model (HMM) is used as a prediction tool to provide optimal QoS. The performance and effectiveness of the proposed PMP approach is evaluated in a real-world test bed and compared with MultiPath TCP (MPTCP), the protocol that is used to aggregate multiple network paths for better network performance. The results show that optimal access point prediction with PMP helps in the WiFi AP selection process compared with MPTCP and conventional approach. By selecting optimal WiFi AP, mobile users experience lower handover count as well as network throughput as compared to MPTCP and the conventional approach.
Keywords
Introduction
The demand of mobile internet today has increased exponentially as more users are relying on it to perform daily tasks such as web browsing, social media access and entertainment. Service providers are tasked to deploy a wide range of 3G/4G cellular networks to cater to the increasing demand despite facing several challenges such as high QoS requirement and cost [1]. As more and more mobile devices are connected to the internet, service providers urgently need a cost-effective method to satisfy the demand for more bandwidth. Despite the promising coverage and bandwidth that can be brought by 3G/4G cellular data networks, WiFi is used by service providers to complement their network service by offloading the ever-increasing mobile traffic especially in urban areas [2]. Customers can be retained when they have better user experience due to higher service capacity and capability. Therefore, WiFi APs are deployed in major cities around the world for mobile users to enjoy seamless access to the internet.
While WiFi offloading can provide better network performance, the dense availability of WiFi APs to cover a broader area and a massive number of users may cause interruption if the architecture is not suitably designed. For example, mobile devices that are constantly on the move may experience frequent and unnecessary handover between different WiFi APs. Furthermore, the lack of intelligence in the device to select the optimal WiFi AP regardless of mobile user movement may lead to service disruption whenever handover occurs. Therefore, one of the commonly used approaches to mitigate the frequent handover issue is to predict the future location of users for a better handover process and efficient resource management [3].
In addition to frequent handover, mobile devices are not designed to select an optimal WiFi AP especially when it is on the move. Optimal WiFi AP connection plays an important role as it will directly impact the quality of experience (QoE) [4]. However, this connection made by the device itself is transparent to users. Therefore, only when users realize their daily tasks are affected by the low performing internet access where QoE is poor, they will manually switch to another WiFi AP with a better signal level [5]. Therefore, mobile devices should be intelligent to select an optimal AP despite constant motion from users. Since the movement of mobile users are far from random and influenced by their historical behavior [6], a prediction based approach shows a promising solution to provide better mobility management.
Aside from predicting user’s future location in order for better handover process and WiFi AP selection, the existence of multiple network paths in mobile devices should be taken into consideration for better QoE. For example, mobile devices nowadays are equipped with both WiFi and cellular network interface. This allows users to use either WiFi or cellular network based on their daily network related tasks and requirement. However, the architecture of mobile devices does not allow the parallel utilization of both WiFi and cellular network interfaces. This prevents users from fully enjoying the benefits of having multiple network interfaces in the device. MPTCP [18], an extension to TCP, was proposed in 2013 that enables the simultaneous usage of multiple network interfaces for better network utilization and performance. This allows mobile device users to gain advantages from both interfaces existing in their devices for higher QoE.
This journal serves as an extension of previous work where hidden Markov model (HMM) is adopted as the progressive mobility prediction (PMP) model to continuously forecast the optimal WiFi AP in future timestamps according to the geographical location of users, hence increasing their QoE [20]. Optimal WiFi AP connection can reduce power consumption because of reduction in WiFi connection to low signal level WiFi AP [7]. In addition, this research compares the performance of the proposed PMP with MPTCP in terms of network performance such as throughput and network interfaces utilization. Previous proposed work of PMP is compared with MPTCP in this journal. The investigation is toward the performance of single optimized WiFi network against multiple networks utilization as proposed in MPTCP.
The rest of this paper is organized as follows: Section II evaluates and presents related works in mobility prediction; Section III presents the proposed system model and Section IV describes the proposed system scheme using HMM as a PMP tool to predict future AP. Testing results of the proposed scheme is discussed in Section V and Section VI outlines the conclusion of the findings.
Related works
Several mobility prediction-based approaches were proposed to predict the future location for mobile users using Markov-based [8, 9] and neural network [10]. Besides mobility prediction, HMM is used in other applications such as human action recognition [11], biology, traffic load prediction, security [12] and rogue AP detection [13]. In this section, we present related works on spatio-temporal and next place prediction, and previous work prior to this research being carried out. Several researches which utilize multiple interfaces for aggregating bandwidth are also presented.
Two prediction models based on HMM namely next place prediction and spatio-temporal prediction have been proposed [14]. As the name suggests, next place prediction is used to estimate the next destination of user if user moves away from current location. Meanwhile, spatio-temporal prediction is used to predict the location of a user at a given time in the near future. In the research, authors first identify point of interest (POI) and convert this to an observable sequence which is used as input to their prediction model followed by user clustering and prediction. Both next place prediction and spatio-temporal prediction are tested in order to find out whether users with a distinct living habit impact the prediction performance. The results show users’ entropy profiles did bring impact on the accuracies of the mobility prediction model and difference among user groups. In specific spatio-temporal prediction, it shows more efficiency to model users with periodical usage compared to next place prediction which is used to model high mobile usage users. The authors would like to highlight that users should always be the one to decide if a prediction model should be applied to the application of their daily lives because of privacy issues.
In a previous work [5], the hidden Markov model (HMM) was used to forecast the next optimal WLAN AP to be connected to in order to improve communication. The proposed approach utilizes the geographical location of mobile users as an observable state and the nearest distance WLAN AP accessed as hidden state in the HMM model. By experimenting the proposed approach in Matlab [15] simulation, the results show that the model successfully forecasts the sequence of hidden states. The hidden states provide the WLAN AP that should be connected to by the mobile user should the user move in the sequence of geographical locations. In addition, the forecast ensures user is always connected to the highest signal WLAN AP, which is described as the optimal WLAN AP. However, this approach leads to a high number of handovers compared to the conventional approach. This is due to the unrealistic test bed that was used in Matlab where WiFi APs are evenly located in 500 m×500 m grid with 30 m between WiFi APs. To deal with this problem, a real-world test bed environment has been setup and an enhanced prediction model is proposed in this research.
The opportunity offered by multiple network interfaces encourage researchers [19] to use Open vSwitch (OVS) to dynamically manage how transmission control protocol (TCP) flows over multiple interfaces that are connected to heterogeneous networks. OVS acts as a bridge to divide incoming traffic from application to be sent over multiple interfaces. More importantly, this process is transparent to application as a virtual Ethernet interface is created to combine multiple interfaces. OVS sends traffic to the virtual interface while the networking stack is responsible for distributing traffic over multiple interfaces. The result shows the ability of seamless migration using multiple network interfaces.
Using multiple interfaces require multipath congestion control algorithms in order to improve resource utilization as well as connection robustness. Dong proposed a new mVeno algorithm where it makes full use of the congestion information of all the subflows belonging to a TCP connection in order to adaptively adjust the transmission rate of each subflow [21]. The performance results in the research demonstrated that the proposed algorithm increases the throughput significantly as well as achieved load balancing compared to existing algorithm.
Progressive mobility prediction
System description
WiFi APs are deployed in dense availability to cover a broad area and a massive number of users. In this paper, the test bed environment consists of N = 5 WiFi AP represented by AP1, AP2, AP3, … AP N is implemented. Mobile users will be connected to these APs according to the prediction results from the proposed dual HMM PMP model. It is also important to note that each WiFi AP’s geographical coordinate represented by (lat,long) together with the mobile coordinate are used in the proposed system. In the test bed, the WiFi APs are placed next to each other in a path with an average of 55 m between each WiFi AP. Meanwhile, all experimental testing will have mobile users initially located near to the first WiFi AP and these users will be walking toward next nearest WiFi AP with average walking velocity of 1 m/s. During the testing, every device will retrieve users’ information containing geo-coordinate as the input to the mobility prediction model every 5 s to perform the prediction. The average span of each testing takes 200 s to finish roaming from first WiFi AP to the last. This experiment is an imitation where mobile users are living a routine life and they are using a similar path going to work hence the devices will be connecting to the chain of WiFi APs along the path.
As described above, mobility scenario is used where a mobile user is entering the range of new WiFi AP j and the previously connected WiFi AP i signal is getting weaker. Traditionally, mobile device tends to use break-before-make approach where the connection with old AP i will only switch to the new AP j if the signal is too weak and break. This process is also known as handover as mobile users switch WiFi AP upon moving. By using dual HMM, each geographical coordinate (lat i , long i ) (hidden state) will be predicted based on previous geographical coordinate with first HMM a . The second HMM b is used to predict the next WiFi AP (AP i ) (hidden state) based on geographical coordinate of mobile user (lat i , long i ) (observable state). Furthermore, all WiFi APs in the described test bed are manually connected beforehand assuming they are available for mobile devices to connect to without further authentication and have sufficient resources to handle every handover process. Therefore, the testing is to validate the proposed approach to forecast the future geo-coordinate of mobile users and predict the optimal WiFi AP to be connected in future timestamp according to the first forecast.
Hidden Markov model
Hidden Markov model (HMM) [16] consists of a finite set of states (hidden states), a finite set of transition probabilities, a sequence of emission variables or output symbols (observable states) and a set of emission probabilities. Transition probabilities denote the probability of a hidden state transition to another state whereas emission probabilities represent the distribution of output symbols that are emitted from each state. Besides, HMM is a doubly stochastic process with an underlying stochastic process that is not observable unless observed through another set of stochastic processes that produce the sequence of observable symbols [17] as shown in Fig. 1.
Example of Hidden Markov Model.
A HMM is represented with a few symbols as below:
H = {H1H2 … H
N
}, H is the N hidden states of the system. In this research, each state H
i
represents WiFi AP
i
geo-coordinate which is the location of the AP. O = {O1O2 … O
N
}, O is the sequence of observable symbols which represents the movement path of mobile users in this research. {π
i
} represents the initial state probability. Initial state probability normally is used to identify the probability of starting at which hidden state H1. In this research, initial state probability represents the starting point of the mobile users. T = {ti,j}, T is the hidden state transition probabilities where ti,j = P (t
k
= H
j
|tk-1 = H
i
) represents the probability of transitioning from hidden state H
i
to hidden state H
j
. In this research, t
i
, j represents the probability of movement from one geo-coordinate to another and the handover from WiFi AP
i
to WiFi AP
j
when i ≠ j. E = {e
i
(k)} where e
i
(k) = P (O
k
|t
k
= H
i
) represents the probability of an observable symbol k emitted from state H
i
. This probability is also known as emission probability.
For ease of use, a HMM can be represented using notation λ = {T, E, π}.
Three common problems can be solved by using HMM.
Problem: Given a HMM model λ = {T, E, π}, calculate P (O|λ). This is known as the likelihood calculation and forward-backward algorithm can be used to calculate the probability of occurrence of the observation sequence O = {O1O2 … O
T
}. Problem: Adjustment of HMM model λ = {T, E, π} to maximize the probability of occurrence of observation sequence (likelihood) and joint probability of observation sequence and state sequence (decoding). Problem: Given an observation sequence O = {O1O2 … O
T
}, decode the hidden state sequence. This process is known as decoding where hidden state sequence H is determined using Viterbi algorithm [17]. In this research, this problem will be focused to predict the future geo-coordinate of mobile user as well as the optimal WiFi AP to be connected by mobile users at specific geo-coordinate and timestamp.
Similar with previous work [5], this research assumes the mobile device can retrieve current geo-coordinate of user in order to perform mobility prediction with the objective to connect to an optimal WiFi AP. Therefore, the geo-coordinate of the mobile user and the location of WiFi AP play a very important role to estimate the parameters hidden in state transition probability as well as emission probability. Regardless, mobile users do not know the exact location of WiFi AP, hence a controller from the operator side which holds the information of WiFi AP is needed. From this perspective, if the controller and mobile device each has the information which is needed to perform the parameters estimation, therefore, the mobile device can request information of WiFi AP from controller to undergo transition probability and emission probability estimation. With this procedure, the device is able to perform the parameters estimation mathematically in mobility prediction and then using the parameters to decode the hidden states (mobile user’s geo-coordinate and optimal WiFi AP selection). Nevertheless, this research aims to forecast on mobile user’s future geo-coordinate and then use the outcome of the first forecast to perform prediction on the optimal WiFi AP to be connected based on the user’s observed geo-coordinate sequence using the Viterbi algorithm.
Progressive mobility prediction using dual hidden Markov model
Operator holds important information such as WiFi AP’s geographical location, while the mobile devices hold their respective geo-coordinate as described in previous section. Together, supervised learning can take place for both transition and emission probabilities whenever the mobile device receives WiFi AP information in order to estimate the optimal WiFi AP as well as future geo-coordinate. However, this is not sufficient as mobile prediction should always be able to forecast a few seconds ahead. For example, where would users go according to their current geo-coordinate and which optimal WiFi AP should the devices connect to if the users were to go to the next location? Existing HMM decoding problem rely on the availability of observation sequence in order to decode the hidden state sequence. However, in reality, mobile user’s geo-coordinate will only be available if mobile user is currently at that location. Thus, this research proposes PMP using dual HMM where forecast can be made with only one mobile user geo-coordinate.
In order to perform the first HMM forecast, mobile user’s first geo-coordinate will be used as the first hidden state in the model. This is followed by placing the first hidden state as the next observable state input to the model. Using this approach, PMP can be made by feeding data into the model whenever the hidden state can be decoded using available observable state. To verify the efficiency of this proposed approach, several experiments will be taken by varying the number of future geo-coordinate prediction the PMP makes. Every future prediction represents the geo-coordinate of 5 s later which is also the duration of the mobile devices when retrieving mobile users’ current geo-coordinate. For example, if two-time step PMP is to be made, the device will only retrieve current user geo-coordinate at the 15th second after the first geo-coordinate retrieval at 5 s and then at 10 s. This is how users’ geo-coordinates are expected to be predicted by the model. After the first HMM prediction, the second HMM forecast which uses predicted geo-coordinate of mobile user as observable state to decode the hidden state, the optimal WiFi AP is predicted.
As depicted in Fig. 2, the first prediction or hidden state, P1 will be decoded by using the previous initial hidden state H1 as the observable state. Then, the following prediction will use the similar approach to produce the progressive prediction.
The illustration of progressive mobility prediction.
Traditional TCP can transfer packet reliably by using connection-oriented service. However, the restriction on communication using single path per transport connection denies the opportunities offered nowadays where multiple paths are available such as WiFi and cellular network in mobile devices. The parallel usage of these multiple paths would improve network resource utilization as well as user experience through higher network throughput and resilience toward network failure.
MPTCP extends TCP by providing the ability to use multipath service where it operates at transport layer and designed to be transparent to both higher and lower layers. Hence, no changes are required by legacy applications to gain advantage from MPTCP. In addition, MPTCP is backwards-compatible with regular TCP where fall back to standard TCP is done automatically without extra actions from user and applications. Figure 3 shows the comparison between MPTCP and regular TCP where regular TCP will use only one network interface during a TCP session while MPTCP support the multiple usage of network interfaces during a MPTCPsession.
The comparison between MPTCP and regular TCP.
In MPTCP, the first packet exchanged between sender and receiver plays an important role to initiate MPTCP connection. This is due to the acknowledgement and additional MPTCP_CAPABLE option in the packet to determine if both parties support MPTCP. If both parties support MPTCP, data transfer will proceed with MPTCP while if either one party does not support MPTCP, regular TCP will be used in that data flow. There are four other possible conditions that can happen in a data flow which are association of new subflow, informing potential address to opposite host, changes in subflow priority and finally connection closes. Association of new subflow happens if either side of the data transfer has multiple network interfaces available to aggregate the connection. Host can also inform potential address to opposite host so that redundant connection takes place if subflow fails. Besides, subflow priority defines the usage of specific subflow to be active or backup. Active subflow transfer packets while backup subflow will only take the role of transferring packet when previously active subflow fails. Lastly, connection close is used to inform the end of a connection.
MPTCP scheduler
The reliability of data transfer in MPTCP depends greatly on the scheduler. There are a total of three scheduler infrastructures namely default, round-robin and redundant which control how packets are transferred in available multiple paths. In default scheduler, data will be sent on the path which has the lowest round-trip time (RTT) until congestion-window is full. Only then, next path with next higher RTT will be used to transfer data. Round-robin scheduler, transmit data in round-robin fashion where multiple paths will be used one after another without preference. The last scheduler namely redundant, transmit data on all available path in a redundant way. This scheduler is useful to achieve lowest latency at the cost of high bandwidth usage.
In the research, MPTCP with default scheduler will be used to test the performance compared with proposed dual HMM PMP in term of throughput and network interfaces utilization.
Performance evaluation
In this section, the performance of the proposed PMP with dual HMM is presented. Optimal WiFi AP prediction will only be made after the prediction of mobile user’s future geo-coordinate. In order to verify the efficiency of the proposed PMP on user geo-coordinate, the distance of the predicted geo-coordinate and the observed geo-coordinate (true value) will be calculated. This process helps to determine how many future geo-coordinates are to be predicted. This method is the most suitable for progressive predicting of user’s future geo-coordinate as well as predicting the optimal WiFi AP. In the experiment, users will roam from the first WiFi AP to the last and is repeated for a different number of future geo-coordinate prediction from one to five.
In order to study the effectiveness of the PMP with dual HMM for network selection, the following metrics were used: -
Distance between prediction geo-coordinate and user’s actual geo-coordinate: This shows the accuracy of the PMP of first HMM. Throughput: This shows the communication quality between predicted WiFi AP and the second HMM. Number of handover occurrence: This indicates the number of transitions between WiFi APs. Multi-interface utilization: This shows the usage of multiple interfaces available in the mobile device.
Figure 4 shows the example of predicted mobile user’s geo-coordinate and actual user’s geo-coordinate. There is a comparison of average distance of predicted geo-coordinates from mobile user’s actual geo-coordinates. Meanwhile Fig. 5 shows the average distance between progressive predicted geo-coordinates and user’s actual geo-coordinates at different numbers of future geo-coordinateprediction.
The illustration of progressive predicted geo-coordinate and user motion path.
The average distance between progressive predicted geo-coordinate and user actual geo-coordinate.
Based on the result as shown in Fig. 5, the highest accuracy of the PMP in the experiment is approximately 8 m of five future geo-coordinates predicted compared to the lowest accuracy of approximately 14 m of five future geo-coordinates. The reason of the average distance of five future geo-coordinates prediction is because the actual path of users cannot be corrected as soon as one future geo-coordinate prediction. For example, if five future geo-coordinates prediction were to be made, the second and above geo-coordinates prediction will rely on the previous geo-coordinate predicted.
If the first geo-coordinate that was predicted has low accuracy, the following prediction accuracy will be lower and lower. On the other hand, if one future geo-coordinate prediction is reasonably accurate, this will ensure the higher accuracy of the prediction by retrieving mobile user’s current geo-coordinate after every future geo-coordinate prediction is made. This clearly indicates that the suitable number of future geo-coordinate prediction the PMP should make is one future geo-coordinate ahead of current geo-coordinate.
In addition to compare the performance of the proposed approach with default behavior of current mobile network architecture, the efficiency of the proposed approach is also compared with MPTCP. MPTCP serves as an extension to TCP by providing the ability to use multiple path service and enabling transport connection to operate across multiple paths between peers simultaneously. With that said, mobile device with both WiFi and 4G network interfaces can be utilized at the same time to increase the overall network performance.
The proposed PMP approach also predicts the optimal WiFi AP based on the first prediction outcome. This helps to select the WiFi AP regardless of where the mobile user heads to. The throughput metric is used in this research to compare with the performance between proposed approach, conventional approach and MPTCP approach where MPTCP utilize both WiFi and 4G.
Based on the real-world test bed, PMP with dual HMM successfully increases the average throughput by 4.21 times compared to conventional approach. It also increases the average throughput by 1.68 times compared to MPTCP approach. The increase in average throughput is understandable as the proposed approach allows higher flexibility where the controller will have the capability to update the optimal WiFi AP geo-coordinate hence allowing the selection of optimal WiFi AP based on mobile users’ geo-coordinate.
Figure 6 shows the comparison of throughput between PMP with dual HMM, conventional approach and MPTCP approach. As shown in Fig. 6, the throughput using PMP with dual HMM is higher than conventional approach in average by 421% and higher than MPTCP approach in average by 168%. The reason for the phenomena is mainly due to the proposed approach successfully assisting the mobile device to connect to optimal WiFi AP for better connection performance compared to the conventional approach where the device is still associated with the poor signal previously connected to WiFi AP instead of handover to a better performing WiFi AP.
The average throughput of progressive prediction using dual HMM, conventional approach and MPTCP.
The proposed progressive mobility prediction optimizes the selection of WiFi AP during mobile users’ movement path. The number of handover metric is important to compare the performance between proposed approach, conventional approach and MPTCP approach.
Based on the results as depicted in Fig. 7, dual HMM PMP yields the lowest number of WiFi handover among the other two approaches namely conventional approach and MPTCP. Dual HMM PMP performed a total of 6.33 WiFi handover in average. The number of WiFi handover performed by both conventional approach and MPTCP are quite similar in average which are 8.67 and 8.33. Hence, the proposed approach successfully decreases the number of handovers by roughly 40% in average.
The number of WiFi handover Occurrence.
The main factor contributing to the decrease of WiFi handover is due to optimization on the WiFi AP selection. The proposed approach dual HMM PMP is able to predict where mobile user is heading to hence it will select the optimal WiFi AP to connect to. What happens in conventional and MPTCP approach is that there is no intelligence involved in WiFi AP selection for both approaches. Mobile device tends to connect to random surrounding WiFi AP during movement. Hence, mobile device may connect to non-optimal WiFi AP where the AP is located far away from where mobile user is heading to hence handover will eventually occur after leaving the coverage of the non-optimal WiFi AP connected previously.
The available network interfaces usually found in mobile devices are WiFi and 4G. However, current network architecture in mobile devices tends to disable 4G network interface whenever mobile devices are connected to WiFi AP. This results in non-optimal utilization of network interfaces that are available to increase the network performance.
Both dual HMM PMP and conventional approaches are not utilizing cellular network when WiFi is connected compared to PMP which utilize network interface 100%. This is mainly due to the current network architecture which prevents the usage of network data which are costly especially for data mobile subscribers where fixed data network quota is allocated for mobile users every month. Therefore, if there is no data network quota concern for mobile users, MPTCP approach shows the most promising results in terms of cellular network utilization compared to the other two approaches. However, by opting MPTCP approach, mobile users may experience a higher usage of power consumption in mobile devices as both cellular and WiFi interfaces are activated to improve the network performance. Besides, MPTCP approach has lower average throughput compared to dual HMM PMP. Figure 8 shows the throughput of WiFi and cellular network interface during downloading a file from internet. Furthermore, the figure also shows that the simultaneously usage of both WiFi and cellular network interface that can aggregate the network throughput at the same timestamp. This suggest that the protocol is able to utilize both network interfaces to complement each other upon one of the interface fails.
The utilization of WiFi and cellular network interfaces in MPTCP.
In this research, a PMP using dual HMM for optimized AP selection is proposed. Dual HMM is utilized as a PMP tool where geo-coordinate plays an important role in predicting the future geo-coordinate of mobile users and the optimal WiFi AP. The PMP is proposed to cope with the limitation of the existing conventional approach where a sequence of observable state is needed in order to decode hidden state.
The results of the first experiment show that the optimal number of future geo-coordinate prediction for PMP is one. In addition, throughput perceived by mobile user increases by four times compared to conventional approach when using the proposed PMP with dual HMM. In conclusion, the proposed approach improves QoS and user experience in the network.
Footnotes
Acknowledgment
The research project is funded by Digi Telecommunications Sdn Bhd under grant No.: 304/PNAV/650764/D113.