An Economical approach for user behaviour determination for the selection of type of Location Based Services

The vision of this research paper is that the mobile phone is aware of its user's motion state and surroundings and modifies its behaviour especially the characteristics of Location-BasedServices based on this information. In the research it is implemented, and evaluated a methodology which can identify individual user states. This learning is expected to occur online and not expected the requirement of any external supervision. The proposed system relies on hidden markov modelling and Log Likelihood Method. The implementation of the methodology is performed by first training the hidden markov model for the required number of states by an intended network trace. The log likelihood value of the data for each hidden markov model in the set is computed and the motion state is identified by choosing the hidden markov model that produced the highest value. The results of simulations carried out in WCDMA environment indicate that the proposed method is able to assist to create a meaningful user context model at various propagation conditions defined by both 3rd Generation Partnership Project (3GPP) and Wireless World Initiative New Radio (WINNER) propagation scenarios while only requiring a network trace without having an integrated sensor onboard cellular phone or any other wearable sensor device.


Introduction
As predicted in [1] there is a bright future for mobile positioning as the key technology for enabling Location Based Services applications, which themselves will become increasingly important as a key enabler of value added services. There are number of methods proposed for determination of user state of a mobile station, however most are based on assistance of an integrated sensor onboard cellular phone or any other wearable sensor device, that is: those methods require modifications to the all applicable mobile transceivers. In [2] Qualcomm Inc., recognizes one of the key requirements for implementation of Location Base Services, is as handset impactminimization of incremental drain of battery power and impact to the handset by new circuitry or software. Some other systems rely on complex algorithms such as Viterbi, hence slower and more burdens on processing. Qualcomm Inc., further states that network efficiency-utilization of minimum over-the-air and backhaul network bandwidth for individual/continuous position reports [2] another major key requirement for implementation of Location Base Services.
Hence it is felt the requirement of proposing a simpler, faster and comparatively lesser processor intensive method with minimum bandwidth consumption for determination of user state of a mobile station at least up to maximum posted speed limits for vehicles and without getting assistance of an integrated sensor onboard cellular phone or any other wearable sensor device, that is: with less modifications to the all applicable mobile transceivers.

Vision, Objective and Scope
The vision is that the mobile phone is aware of its user's state and surroundings and modifies its behaviour especially the characteristics of Location-Based-Services based on this information.
The objective of this research is to propose a simpler, faster and comparatively lesser computationally burdened method for determination of user state of a mobile station up to maximum posted speed limits for vehicles 1 and without getting assistance of an integrated sensor onboard cellular phone or any other wearable sensor device, that is: with less modifications to the all applicable mobile transceivers to WCDMA 2 .
In this paper it a relatively economical system which can identify individual user states will be proposed, implemented, and evaluated. This learning is expected to occur online and does not require any external supervision. Proposed method is not intended to be a total solution; however the process based on the proposed method is intended to be a part of an organization of services.

Survey of Previous work
This section summarizes current knowledge and what other authors have done in the various topics within the purview of this research paper.
In [3], it has discussed two different filtering approaches: namely Hidden Markov Model and Kalman filter, in order to minimize the impact due to ambiguity and signal degradation, on position estimates in cellular positioning based on database comparison method. He points out that Hidden Markov Model based methods have fewer parameters and easier to estimate, the estimator seems to be less sensitive to errors in the model parameters and multi hypothesis processing is simpler. He concludes stating that the Hidden Markov Model based approach is more suited to perform position estimation of a database comparison process.  [5], a design, implementation, and evaluation of a wearable system which can learn contextdependent personal preferences by identifying individual user states and observing how the user interacts with the system in these states.
Ian Anderson and Henk Muller [6] have showed that by using patterns of signal strength fluctuations and changes to the current serving cell and monitoring neibhouring cells in GSM, it is possible to distinguish between various states of movement using Hidden Markov Model compares with the results of method based on Artificial Neural Network. However this paper utilizes a relatively complex Viterbi algorithm to infer most likely state sequence.

Overview of this Paper
Next chapter covers theoretical background, followed by an introduction to the proposed method and simulation methodology respectively at chapters 4 and 5. Chapter 6, presents results and relevant analysis. The paper is concluded with chapter 7, which covers summary, conclusion and proposed future work.

Markov Property
If we know the present state, then knowing past states gives no added information about the future states. Given the present, the future is conditionally independent of the past.
where 5 is the set of all states, and 5 is the state variable.
When using Markov models for modelling a physical phenomenon, we generally only have Maximum   In addition to the Markov property between states S., this model makes a further assumption that the observations are conditionally independent given the current state.
As above this implies that state 5 captures all relevant information of the system and it further implies that any measurement is independent given the state. Estimating the state 5 ; . from observation O, a known observation model is described as an inference or state estimation. Using data to find the transition and observation models are generally called learning or parameter estimation.

N-state Hidden Markov Model
The time increment is equal to the bit duration.

Learning of Hidden Markov Model
In view of improvement of initial model parameters F=(n, A, B), the expectationmaximization (EM) algorithm is utilized. Here the complete log likelihood Pr(<9 , S|T), that a sequence O, of data bits C^were generated by the hidden markov model is maximized by adjusting the hidden markov model parameters T. Since of incomplete data, because of the states 5 are hidden, the expected complete log likelihood should be maximized, with respect to the parameters F. The maximum likelihood parameters are the improved hidden markov model parameters, and the process consisted of the following steps: StepO
Step 2 Computation of the state transition probabilities given Oand Fand update the parameters A, and B according to: Step3

Recognition of State
At each successive time step it is needed only to update a by a.
That is sum for a (i) over all possible states 5 multiplied by the transition probabilities of (9) shifting from that end state to the current state 5 and then multiplying by the probability of observing the current observation O. from state 5. The end result will be the likelihood of the data for the particular hidden markov model.

Assumptions
Further following were assumed for the simulation: Firstly that: the base station employs an idealized rake receiver. The rake receiver is idealized in the sense that: the number of taps in this receiver is equal to the number of multipath components and the receiver has perfect knowledge of the delay of each multipath component and the rake receiver uses the pilot symbols in the DPCCH 3 to estimate the phase offset generated by the channel. Then, the receiver extracts each multipath component and employs an equal gain combiner to generate a test statistic, which is evaluated via a simple zero -threshold test. Secondly it is assumed that each component of the multipath channel has the same Doppler spread.

Simulation of Physical Layer of 3G Wireless System
The simulation 4 was performed on a Pentium™ D CPU 3.40GHz based personal computer with 1024 MB RAM and Matlab™ R2006a (version 7.2.0.232) was utilized as the simulation platform.
WCDMAsim(TM) was utilized as the simulator of physical layer of 3G wireless system. The WCDMAsim 5 is a Matlab™ based, physical layer simulator of a 3G wireless system that conforms to the WCDMA standard. This simulator emulates the behavior of WCDMA base and mobile stations in specular multipath fading environments. WCDMAsim provides a technically accurate and detailed emulation of a (14) WCDMA system [7]. During the simulations the uplink simulator emulates the operation of a WCDMA mobile station in a frequency selective, multipath fading channel. The simulator emulates up to six DPDCH s s and the DPCCH 7 . However here the common physical channels (such as the Random Access Channel and the Physical Common Packet Channel) and the channel coding (specified in ETSI TS 125.212) are not modelled. The simulator output is an error length sequence. The error length sequence contains information about the number of consecutive symbols that were received errorfree and the number of consecutive symbols that were erroneously received.

Simulation of Propagation conditions
The standards proposed in [8] and [9] by the 3GPP for radio transmission and reception was utilized for the simulation and there are number of different multipath channel propagation conditions for which the proposed methodology could be simulated as follows:             Finally here the fourth row provides the maximum value of log likelihood for 120 bit length obtained from a mobile station (which was moving at 130 kmh), incorrectly recognizing that the mobile is at the state of moving at the nearest speed of 70 kmh.

Summary
The results indicate that the proposed method is able to assist to create a meaningful user context model for the selection of the characteristics of Location Based Services, at various propagation conditions defined by both 3GPP and WINNER propagation scenarios (up to a maximum speed of 120 kmh) as a passive method only requiring a network trace, and without an integrated sensor onboard cellular phone or any other wearable sensor device, that is: with less modifications to the all applicable mobile transceivers to WCDMA.

Evaluation
The selected statistical HMM models are generally very rich in mathematical structure and hence can form the theoretical basis for use in a wider range of applications. The method of maximum likelihood provided estimators that have both a reasonable intuitive basis and many 11. Bit Error Rate.
desirable statistical properties. The main reason for the selection of maximum likelihood method is that it is very broadly applicable and simple to apply. Identification of the state was simple as: computation of log likelihood of the data for each hidden markov model in the set and identify the state by choosing the hidden markov model that produced the highest value without utilization of complex algorithms such as Viterbi, hence significantly faster and lesser burden on processing.
For the verification of proposed method the WCDMAsim™ was utilized as the simulator of physical layer of 3G wireless system developed by Mobile and Portable Radio Research Group at, Virginia Polytechnic Institute and State University, Blacksburg, Virginia, which provided a technically accurate and detailed emulation of a WCDMA system.
One unique feature of WCDMAsim is its ability to produce more than just average BER 11 statistics. Rather, this simulator provides users with an error length sequence that contains not only the number of errors but also the location of each error in the data sequence. This information is crucial for the development of computationally efficient simulations, such as Hidden Markov Model simulations of largescale wireless networks [30].
The proposed method found to advantageous in view of that it eliminates integrated sensor on cellular phone and hence minimum modification to the existing of future WCDMA compatible mobile phones.
Further the proposed method eliminates any sort of signalling of data on the user state from Mobile station to base station or vice versa, since the proposed method is a one way passive process of analysing a received network trace.
Generally for state recognition methods based on channel estimation methods: faster the mobile station is, shorter the optimum pilot length for channel estimation will be, due to the difficulty in tracking faster channel variations, hence lesser accuracy. This minus point will become a plus point for the proposed method, since it is taken number of error symbols with compared to number correct bits into account during recognition of user state.
Further to above higher the mobile station speed, more frequently the signal fades, again the proposed method gains the advantage as explained above. The automatic or unsupervised initiation of the process could be performed by sensing the change of Cell ID or could be incorporated to any other mobile location method utilized.
Jeffrey H. Reed et al [13] have recognized first major challenge as radio frequency channel imperfections particularly multipath time dispersion and interference. The proposed method prevents most of problems caused by NLOS and multipath propagation (which is already taken into account in the network trace), where as the accuracy of channel estimation methods, dependent upon the measurements, degrade significantly.
Finally a large amount of data was involved in analysing the output of the WCDMA simulator and consequent conversion of same to a bit train of '1's and 'O's, which was the input to the building of hidden markov models: that is improvement of initial model parameters as explained in above paragraph 4.1. It could have been better if any data clustering method such as K-means or Principle Component Analysis was utilized in order to extract significant components of data.

Applications
The process based on proposed method could be a part of organization of services and a single service interface that provides applications with contextual information. Such a system is proposed by Glenn Judd and Peter Steenkiste in [14] which introduces a concise organization of services and a single service interface that provide applications with contextual information in a unified manner and shows, via example applications and services that have been implemented, how the proposed service organization and interface can be used to allow proactive applications to adapt their behaviour to match a user's current environment.
V Bellocci et al [15] identify location-based services: News, Traffic Information, Yellow Pages Services and Fleet Management are as services that require low level of accuracy. For such types of services, utilization of a complex and expensive method to determine user behaviour is not worth and the proposed method is expected to be worthwhile for implementation for such services due to its simplicity and comparatively low cost compared to hardware based systems. Juh.i Korhonen [16, p474], emphasizes Unimportance of cost criteria in the selection of « service by a future 4G user (or rather a smart device) as, 'The smart device can analyze its environment and choose the best available service provider and technology (service discovery). The evaluation is based not only on radio path characteristics, as in current networks, but on other parameters, such as QoS 13 , cost, and security', unquote.
Monitoring the speed in passive mode of a particular mobile user, would be very convenient for law enforcement agencies as there would be no need for special speed radars as well as it may also be used to activate a camera which is in sleeping mode, or they could send an instant speeding ticket via 3G device.
Finally even though the speed of the user was selected as the parameter of recognition, the same method may be utilized to recognize a user by utilization of any of the other appropriate parameters. 12

References
The computational burden associated with the Baum-Welch algorithm is quite higher for longer error vectors, since the forward and backward variables are computed for each symbol and convergence is observed to be slower. Sivaprakasam and Shanmugan [17] have showed that under some assumptions, a markov model with an arbitrary transition matrix is equivalent to a markov model with a unique block diagonal transition matrix and presented a computationally efficient algorithm, for estimation of transition matrix from a set of observations, called modified Baum-Welch algorithm. However they have shown that the particular algorithm is computationally efficient when the overall bit error rate for the source is small and the error sequence contains long stretches of identical symbols [17]. The utilization of modified Baum-Welch algorithm might accelerate the real time execution of proposed methodology.
Further to above the conclusions by Andrea J. Goldsmith and Larry J. Greenstein in [18] states that the short term fluctuation statistics of signal attenuation, for LOS(NLOS) regions of a microcell are Rice-like(Rayleigh-like), may be utilized as priori information in further improvement of the proposed method. For example generation of markov model parameters such as initial state matrix could be based on either of Rice-like(Rayleigh-like), models as per the situation.
Finally let us conclude quoting K W Richardson [19], who states: 'the range of 3G services will truly only be limited by the imagination of the service provider', unquote.