A Situation-Aware Approach for Dealing with Uncertain ... - IEEE Xplore

A Situation-Aware Approach for Dealing with Uncertain Context-Aware Paradigm Xiangtao Lin1 , Bo Cheng2 , Junliang Chen3 State Key Laboratory of Networking and Switching, Beijing University of Posts and Telecommunications, Beijing, China, 100876 1 [email protected], {2 chengbo, 3 chjl}@bupt.edu.cn

Abstract—Context-aware paradigm is intended to make decisions proactively for users to adapt to contexts changes in a pervasive environment so as to improve users’ work efficiency. However, this commitment deduces because of the intrinsic uncertainty i.e. incompleteness, inaccuracy and inconsistency of context-aware paradigm. We bring forward a situation-aware approach, supported by Bayesian Networks, ontology and Domain Specific Language techniques, for dealing with uncertain contextaware paradigm. Situation, a description of logically combined contexts, is high-level abstract contexts; hence, it can shield the trivialness and inconstancy of low-level contexts. BN mapping contexts to situations is good at dealing with incomplete, inaccurate and erroneous low-level contexts; ontology is referred to eliminate inconsistencies among situations and contexts. Besides, DSL can offer flexibilities for its easy readability and good reusability. Also, interruption ratio, precision and efficiency are evaluated to validate the effectiveness of our approach w.r.t. a situation-aware multimedia conference application. Index Terms—situation-aware, context-aware, uncertainty.

I. I NTRODUCTION ontext-aware paradigm is intended to make computing devices invisible from users’ attentions so that users can focus on tasks themselves instead of computing devices. In context-aware paradigm, a user may have various kinds of devices like laptop, PDA and intelligence handsets at the same time; in addition, the environment is flooded with various kinds of sensors from physical sensors, detecting physical environment factors e.g. temporal-spatial information, temperature, humidity etc., to virtual sensors, collecting application’s runtime parameters like workloads, available memories, QoS statistics and so on. Moreover, various kinds of services covering different networks from internet to telecom network are deployed in pervasive computing environment. The characteristics or factors about these entities e.g. users, devices and services etc. are considered as contexts [1]. Under these circumstances, a user will be exhausted to deal with those devices and services since they have to interact with these devices or services frequently with regard to their contexts changes. To complicated matters, these contexts about different entities from diverse domains may differ in essential and are distinct from data structure which require special processing. Contextaware computing [1] paradigm was put forward to facilitate

C

Our work is supported by ”973” program of National Basic Research Program of China under Grant 2007CB307103 and National Natural Science Foundation of China under Grant 60432010.

contexts acquisition, representation, aggregation, interpretation and utilization [2]. These facilities extract the semantics of contexts from different domains, simplify personalized services creation, decrease the frequency of interactions between users and computing environment by making decisions for users according to current contexts and context changes, and enhance the efficiency of pervasive systems. However, in many cases, contexts are low-level or even direct abstractions of raw data from various kinds of sensors planted in ever-changing environments; as a result, contexts are trivial, inconstant and multiform. Besides, contexts are uncertain for their incompleteness, inconsistence and inaccuracy since contexts reflect the real-time statistics about uncertain environment. The uncertainty characteristic [3] of contexts is proved to be a big obstacle on the way to context-aware paradigm. For example, a designer or a user will have difficulty in deciding to pick up which action to be the next step in a context-aware paradigm, when some contexts are partially distorted or even missing due to network overload or sensor failure. In this paper, a situation-aware approach is employed to overcome the problems brought by trivial and uncertain contexts in context-aware applications. In short, a situation [3] is considered as logically aggregated but more abstract contexts. In our situation-aware approach, applications will adapt their actions to situations instead of contexts solely, which will overcome the trivialness and uncertainty problems. We will show the situation-aware approach can assure less user interruptions, good precision and higher efficiency. The reminder of this paper is arranged as follows: in section 2, we will argue the reasons for which situation-awareness is employed in our work. Our situation-aware approach is addressed in section 3. In section 4, a situation-aware multimedia conference application is described and explained. Section 5 evaluates the performance of our approach. In the end, we will make a conclusion and discuss the future work. II. F ROM C ONTEXT-AWARENESS TO S ITUATION -AWARENESS A. Pitfalls of Context and Context-Awareness Although context-aware paradigm provides users with more pervasivenesses [4] i.e. less interruptions, and higher work efficiency comparing with traditional applications, it still has some

978-1-4244-4148-8/09/$25.00 ©2009 This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE "GLOBECOM" 2009 proceedings.

B. Highlights of Situation and Situation-Awareness What brings context-aware paradigm to that embarrassing point is that context-aware paradigm is fine-grained and in low-level abstraction. As a result, too many factors have to be taken into consideration; and, applications have to adapt to any changes of those factors and halt for any uncertainty of those factors. Naturally, when we reconstruct contexts and contextawareness in higher-level abstraction, we will get situations and situation-awareness.

Definitions of Situations and Situation-Awareness Definition 1 A situation is a description [5] of logically combined contexts during a period and it is a higher-level abstraction of these contexts. Figure 1 depicts the abstraction relationship between situations and contexts. We can see that situations are at a higher level than contexts; hence, situations are less than contexts and the relationships between situations are not as much complex as that of contexts. So situation-aware computing seems to require fewer resources and be more efficient than contextaware computing. High Abstraction level Low

disadvantages due to the intrinsic characteristics of contexts and contexts-aware computing paradigm. In general, contexts in a specific domain have the following characteristics: Contexts Are Trivial Contexts usually come from various kinds of sensors which may vary a lot in essential and domains. Those sensors can be physical sensors e.g. RFID readers or software sensors like workload monitor. The information collected by those sensors is of different type of raw data in diverse structures. It is a challenge for engineers to fuse these raw data into a unified representation form and mine useful contexts to make right decisions to guide the next step for a system. Contexts Are Inconstant Most of contexts, except for user profiles and so forth, describe the characteristics of ever-changing environment and reflect the real-time statistics of uncertain environment. Inconstancy presents a fact that sensor readings often change frequently in order to catch up with the dynamic and complex environment changes. The semantics of sensor readings i.e. contexts often change with passing time, shifting places and transiting application states. Contexts Are Uncertain Uncertainty is thought to be the most essential characteristic of contexts. Uncertainty is interpreted as incompleteness, inaccuracy and inconsistency. Incompleteness, usually caused by network overload or sensor failure, is occurred if the expected contexts are missing. Inaccuracy, often caused by imprecise sensors and/or distorted raw data, is happened when the detected data falls out of the right form or precision range it is supposed to be. Inconsistency, often caused by redundant contexts, is about the contradictory of contexts coming from different sensors at a certain time. Obviously, applications have to adapt frequently to finegrained, trivial and inconstant contexts to catch up with dynamic environment changes. Computer cycles are wasted heavily since any trivial contexts changes will lead to states transitions which may be unnecessary. Uncertain contexts will bring applications to frequent interruptions. For instance, applications will make no headway with incomplete contexts; they will be at a loss with inaccurate contexts and they will halt between two options with inconsistent contexts. Under those circumstances, applications can not make decisions with uncertain contexts and have to ask users for further instructions; hence, users are notified and interrupted and application efficiency is brought down.

traveling resting

C1 C2

Situation Layer

working

C5 C4

C3

Context Layer

Sensor Layer

Fig. 1.

the abstraction relationship of situations and contexts

Definition 2 Situation-awareness [1] is the ability of using situations to provide domain specific computing abilities to users. In our work, a multimedia conference application adapts its actions to situation transitions. This procedure is considered as a kind of situation-awareness by definition 2. Advantages of Situation and Situation-Awareness Situations describe the states of the entities of a system; they come of contexts and are regarded as semantic interpretations of a collection of contexts [6]. They are derived from lower level contexts while the drawbacks of contexts, the trivialness and uncertainty problems, are intended to be eliminated. Situations are at a higher abstract level and they are much less than contexts defining them; thus, they have a less number of data forms or structures. Therefore, the trivialness problem of contexts is weakened or even eliminated. In addition, situations are much more stable or even constant than their resources i.e. lower-level contexts. A situation is alternated unless several of its referring contexts are changed. Only one change may not lead to situation transition. For example, John is on his way to office by car connecting to a multimedia conference through his cell phone; and, he belongs to a member meeting situation and the changes of his location (context) will not foreclose him out of the member meeting situation. Situations and situation-awareness are appropriate to deal with the uncertainty problem of contexts. We will show how this claim is guaranteed in detail in next section. III. A S ITUATION -AWARE A PPROACH A. Dependencies among Situations and Contexts Figure 1 merely depicts the abstraction level between situations and contexts without any further details. From the definition of situation in previous section, we can see that a situation often depends on several contexts and conversely a context may be involved in several situations. In fact, causal dependency is regarded as the most important relationship


among situations and contexts. By definition 1, a situation is said to be satisfied only when all of its depended contexts are available, fall into their right ranges and have their appropriate values. Obviously, uncertainty problem remains unresolved until now. There are many approaches for dealing with uncertainty [7], for example, Dempster-Shafer theory, Three-Valued Logic, Fuzzy Logic and Bayesian Network. Bayesian Network (BN) is highlighted since it guarantees fairness to involved entities. BNs are theories to deal with uncertainties based on probabilities. In artificial intelligence, BNs are usually used to calculate the probabilities for decision making to deal with uncertainty problems. A BN is a direct acyclic graph whose nodes represent discrete or continuous variables and arcs are causal dependencies between connected nodes. Node A is called a parent of node B and B is a child of A if a direct arc from A to B is present. The arc implies B depends on A directly. A node is called root node if it has no parent and depends on nothing. Each root is associated with a prior probability specified by experts. Each of other nodes is associated with a Conditional Probability Table (CPT) which records all possible combinations of the values of all its parents when the variable represented by that node is discrete. In our approach, Situation-Context Bayesian Network (SCBN) is employed to deal with uncertainty problems of context-aware paradigm. In a SCBN, situations are the most junior descendants in the SCBN family. The other nodes represent contexts since contexts are abstractions of sensor readings and depend on sensors only. A situation depends on all of its parent nodes e.g. contexts. The causal dependencies among situations and contexts are captured qualitatively by arcs connecting them and the strength of these causal dependencies is quantitatively encoded by CPTs associated with situation or context nodes [8]. So far, we give a situation decision theorem w.r.t. SCBN: Theorem 1 A situation s is satisfied or occurred if the probability p of its corresponding node x in SCBN is no less than a threshold t , where

above. As sensors are independent with each other, root variables or nodes of SCBNs are conditional independent. Further more, we have conditional independencies assumption of node x and all other nodes given parents(x) that is node x only depends on their parent nodes. So the CPT of node x only contains conditional probabilities given parents(x), which can save one half of storage since we only need exponential storage in the number of parents(x).

p = P (x = s|assignments of parents(x))

A similar procedure is employed to handle inaccuracy problem except that the missing contexts are substituted with inaccurate contexts since the latter can be regarded as missing contexts for their uselessness. The above procedure is easy to be carried out and will gain good efficiency because the structure of SCBN needs not to be rebuilt during the procedure. Ontology techniques are referred to deal with inconsistency problem, which is usually caused by redundant contexts contradicting in semantics. A Situations-Contexts Ontology (SCO) is employed to model concepts and other relationships, except for causal dependencies relationship, among situations and contexts. Figure 3 shows the hierarchy of a situation ontology fragment, which is used in our multimedia conference application. When inconsistency is occurred that is applications are halt between two options with contradicting situations or contexts, the ontology is consulted to help eliminate the inconsistency as long as they are similar in semantics w.r.t. similarity degree.

= P (x = s|x1 = c1 , · · · , xn = cn )

(1)

, x1 , x2 , · · · , xn are parent nodes of x, whose values are assigned with contexts c1 , c2 , · · · , cn , respectively; and, s is the situation node x represents. To make things simplified and get high efficiency, variables of a SCBN are made discrete and their ranges are made TRUE or FALSE as much as possible. They can be reached by sampling continuous values into discrete samples or/and converting situations or contexts into predicates. For instance, A PDA may have three modes (contexts) i.e. Sleeping, Meeting, Driving, and we prefer to add three nodes, mode(PDA, Sleeping), mode(PDA, Meeting), mode(PDA, Driving) , to SCBN rather than one variable context(PDA, Mode) since the latter node is a three-value variable while the former three nodes are two-value variables i.e. TRUE and FALSE. In our approach we trade space for speed in the manner mentioned

B. Dealing with Uncertainty With BN learning techniques, incompleteness problem can be handled by filling in the blanks. We fill in missing contexts with aggregated values like mean values or statistical values [9] of historical contexts. Besides, quantitative causal dependencies carried by CPTs associated with missing contexts are also used to compensate for incompleteness. To capture quantitative causal dependencies, Expectation-Maximization algorithm [10] is applied to training set with steps in figure 2, in which step 1 and step 2 are Expectation stage and the other steps are Maximization stage. Begin

①

historical data of missing contexts

Calculate Expectations of missing contexts

②

Fill in missing contexts with Expectations

③

Likelihood functions about missing contexts

End

Find new values of expectations using Maximum-LikelihoodEstimation technique Y

？

Convergent N

④

Substitute Exceptions with these new values

Fig. 2.

the procedure dealing with incompleteness


C. Building Situation-Context Bayesian Networks

15 #enable each service of serviceList

In this subsection, we give an algorithm building a SCBN. Before that, the ontology SCO mentioned in subsection B is created to model concepts about situations and contexts. In this ontology, only generalization-specification relationships among concepts are described in order to decrease the complexity of similarity calculation and gain higher efficiency. In addition, in the algorithm, we apply native Beyes ideology [11] to calculate the joint probability distribution in line 18. We assume that an uncertain node (context) is caused by its least 3 ancestor nodes, which has the same theory as 3-gram proved to be efficient by Google n-gram information extractor [12].

16 [RHS] (serviceList) should be enabled 17 = FOR each s IN ”serviceList” 18 enable(s)

1: INPUT: situation definitions, SCO 2: OUTPUT: a SCBN 3: BEGIN 4: FOR each situation definition S DO 5: Split S into normalized context concepts Cs

9: 10: 11:

situation

resting

by referring to SCO; Choose an order for Cs and the concepts mapping to sensor readings are moved up to the front;

Fig. 3.

Add node vi to SCBN; Set to be the minimal subset of v1 , v2 , · · · , vn−1 such that vi is conditionally

15: 16: 17:

independent with all other members of v1 , v2 , · · · , vn−1 given parents(vi ); Give the probability table of P (vi |assignments of parents(vi ));

D. Defining Domain Specific Language In situation-aware applications, when a situation is identified, reactions should be taken. In our approach, a Domain Specific Language (DSL) is defined from a specific rule language Drools [13] to make problems of a specific domain to be expressed clearly and easily and to gain well-defined modularity and reusability. An example DSL statement and its definition are given as follows: 1 #if both workingAtCompany and businessMeeting 2 #situations are satisfied, then services in serviceList 3 #should be enabled

8 #the definition of Left Hand Side of the above 9 #DSL statement in rule engine specific language 10 [LHS] Situation(x) AND Situation(y) 11 = satisfy (id ==”{x}”) == TRUE 12 satisfy (id == ”{y}”) == TRUE 13 #the definition of Right Hand Side of the above 14 #DSL statement in rule engine specific language

onBreak

dining

onLeave

walking

working

onBus

driving

working at home

inTaxi

working at company meeting

member meeting

12: 13: 14:

4 IF Situation(x) AND Situation(y) 6 THEN (serviceList) should be enabled 7 #the interpretation about the above statement

sleeping

traveling

compatible

Convert concepts into predicate variables v1 , v2 , · · · , vn (see subsection A); FOR each variable vi DO

18: 19:END

The situation-aware multimedia conference application provides the members of a company or a group with a multimedia conference service, to which the members can access whenever and wherever they need. It adapts to situations instead of contexts; that is, when contexts change, a situation is identified by theorem 1 and reactions w.r.t. this situation in DSL is carried out. To be concrete, the application will do nothing if contexts changes are not strong enough to lead to situation transitions; otherwise, the reactions are issued to adapt to situation transitions.

compatible

6: 7: 8:

IV. A S ITUATION -AWARE M ULTIMEDIA C ONFERENCE A PPLICATION

business meeting

situation hierarchy of multimedia conference application

Figure 3 shows the situation hierarchy of the application. The compatible situations can be occurred at the same time. Next, we describe situation transitions and their corresponding reactions through a scenario. John prefers to work at home rather than company as he lives far from his company. Now, contexts reveal that he is involved in workingAtHome and memberMeeting situations. He can see, hear and talk to his colleagues by his PC, microphone and camera through multimedia conference application. In the middle of the meeting, he is asked to come to the company to meet a business partner since the partner insists to meet with John, who is the technique manager. On his way to his car, he is in memberMeeting and walking situations since his colleagues will tell him information about the partner during his way to company. He is accessed to the application by his cell phone. When he drives to company, he is involved in driving and memberMeeting situations. He is connected to the application by PDA residing his car. A reverse procedure is happened when he gets out of his car and walks to his office. Once he meets with the partner, he is in workingAtCompany and businessMeeting situations. He is accessed to the application by his PDA. Services and devices enabled in every situation are showed in figure 4. For example, w.r.t. the rightmost vertical channel, when John is in workingAtCompany and businessMeeting situation, he is connected to the application by PDA and he prefers to receive all important calls, emails and Instant Messages. V. E XPERIMENTAL VALIDATION Three experiments to be addressed in this section are carried out to verify the effectiveness of our approach w.r.t. the


in order to reach a statistical trend.

Context- Situationaware aware paradigm paradigm

time Situation

BN

context

memberMeeting workingAtHome walking

device

computer

service

DSL

profiles

all calls all emails

cell phone

driving PDA

businessMeeting walking workingAtCompany cell phone

important important important calls calls calls important emails

all IMs

Fig. 4.

PDA important calls important emails important IMs

a situation-aware scenario

situation-aware multimedia conference scenario introduced in previous section. There are 3 kinds of devices, 6 types of services and 5 situations in this scenario, which is showed in figure 4. We refer to 10 kinds of contexts, which are location, RFID, speed, schedule, PDA state, PC on/off, network status, camera on/off, microphone on/off and users’ preferences and so on, to define situations or make decisions in context-aware paradigm. In our experiments, the application decides which device to access or which service to invoke according to situations or contexts changes. A. Decreasing Interruption Ratio When contexts and context-awareness are introduced into traditional applications, they are intended to help users to react proactively to various entities of pervasive systems so that users are not interrupted frequently and focus on tasks they are dealing with instead of those entities. Interruption ratio is a ratio of the number of interruptions to the number of decisions adapting to contexts or situations changes. As Context-aware applications can not fulfill all their commitments because of their intrinsic uncertainty characteristic and our approach has claimed to conquer uncertainty problem so as to decrease interruptions, interruption ratio of our approach should be compared with that of context-aware paradigm to verify whether observations meet the claims and the degree of the decrease if the decrease does exist. We define interruption ratio as following: notif ying w3 N interruption ratio = active +w2 N passive +w3 N notif ying w1 N

(2)

active represents the number of observations that , where N the situation-aware application decides to take actions actively passive represents the number of observations for users, N that the situation-aware application decides to take no acnotif ying represents the number of observations that a tion, N situation-aware application can not make a active or passive decision itself and notifies users for further instructions that is interruption occurs. w1 , w2 , w3 are weights of the above three factors and they can be tuned to adapt to different scenarios. In general, they hold w1 + w2 + w3 = 1. In our experiment, notif ying contributes interruption the most, N passive we think N active the last so that we have w1 = 2/9, the second and N w2 = 1/9, w3 = 6/9. We ran the experiment about 700 rounds

Fig. 5.

interruption ratios of CA-MCA and SA-MCA

In figure 5, triangles are interruption ratios of context-aware Multimedia Conference Application (CA-MCA) and squares are interruption ratios of our situation-aware MCA (SA-MCA). Dotted line and solid line are polynomial fits of interruption ratios of CA-MCA and SA-MCA, respectively. We can see that SA-MCA gains less interruption ratio than CA-MCA in most cases. That is because the total number of decisions is almost immutable in a specific scenario i.e. MCA; and, our situationnotif ying which in turn will aware approach can decrease N active and N passive with a same number as decreased increase N notif ying ; as a result, the interruption ratio of our SA-MCA N will decrease and users’ experience will increase. B. Raising Precision Our situation-aware approach decreases interruption ratio by eliminating uncertainty of context-aware paradigm. Many vague decisions that are supposed to notify users become decidable and transform to active or passive ones. The precision of these redistributive decisions should be evaluated since wrong decisions will distort users’ preferences and pull users’ experience down. The precision is defined as follows: precision = w1

notif ying passive active N N N + w2 + w3 Nactive Npassive Nnotif ying

(3)

, where w1 , w2 , w3 are the same as those of interruption active , N passive and N notif ying are observations interratio;N preted in subsection A and Nactive , Npassive and Nnotif ying are expectations of these observations, respectively. Usually, these expectations are derived from historical statistics of traditional paradigm without context-awareness. Figure 6 shows precisions of CA-MCA in dotted line and SA-MCA in solid line. The precision of SA-MCA is below notif ying but close to the precision of CA-MCA although N of SA-MCA is less than that of CA-MCA. That is because MCA is not a notifying intensive scenario that is Nactive and Npassive are the majority of the total number of decisions so active /Nactive and N passive /Npassive are that the increase of N notif ying /Nnotif ying as the latter slower than the decrease of N has a less denominator i.e. Nnotif ying . With notifying intensive applications, our approach will be more precise than contextaware paradigm according to precision definition.


uncertainty referring to BN and ontology, it is traded for denotif ying greatly so that Taccept and Treject are saved creasing N in our approach. For that reason, by figure 7, our approach outperforms for about 30% over context-aware paradigm in efficiency. In addition, our approach has an additional time consumption of building BNs for an application and that time consumption is not count into efficiency since the building procedure runs only once for a scenario and after that time consumption does not effect efficiency any longer.

Fig. 6.

precisions of CA-MCA and SA-MCA

C. Increasing efficiency Context-aware paradigm aims to improve work efficiency by making decisions proactively for users according to contexts changes. So the efficiency of our approach should be assessed to insure whether our approach is meaningful. The efficiency is defined as following: eff iciency = w1 Tactive + w2 Tpassive + w3 Tnotif ying

(4)

, where w1 + w2 + w3 = 1, Tpassive = Tdecide , Tactive = Tdecide + Tservice , Tnotif ying = Tdecide + p(Taccept + Tservice ) + (1 − p)Treject , 0 < p < 1, and Tactive is the average time of deciding to invoke service actively for users. Tpassive is the average time of deciding to take no actions for users. Tnotif ying is the

average time of deciding to notify user and taking actions w.r.t. user’s response i.e. acceptance or rejection. Tdecide is the average time of making a decision in a scenario. Tservice is the average time of a service execution. Taccept and Treject are the average responding time for users to accept or reject the suggestions a notifying gives, respectively. We assume users will answer a notifying once they receive it; then, we will get Taccept = Treject and Tnotif ying = Tdecide + pTservice + Treject . In our experiment we assume p = 0.5. Figure 7 gives the normalized efficiencies of CA-MCA and SA-MCA.

Fig. 7.

efficiencies of CA-MCA and SA-MCA

Our approach decreases interruptions without precision degradation. Although Tdecide of our approach is a little bit of more than that of context-aware paradigm for dealing with

VI. C ONCLUSION AND F UTURE W ORK In this paper we have presented a situation-aware approach for dealing with uncertain context-aware paradigm. In summary, our approach deals with uncertain context-aware paradigm at a higher abstract level i.e. situation-awareness instead of context-awareness. We follow a modeling-reasoningdecision making process. Specifically, BNs and ontology are referred for situation modeling and reasoning, by which uncertainty is eliminated. After that, DSL is employed to express decisions w.r.t. a situation. We give a multimedia conference to show how situation-awareness paradigm works. Our experimental validations prove that situation-aware paradigm outperforms uncertain context-aware paradigm in three aspects i.e. interruption ratio, precision and efficiency. In future, we will focus on the automation of the whole process. We will consider autonomous situation learning from low-level contexts and now situations are specified by experts. The boundary between situations and contexts should be autonomously figure out definitely. We also will introduce feedback mechanism into decision making. R EFERENCES [1] A. Dey, ”Understanding and Using Context”, Personal and Ubiquitous Computing, vol. 5, no. 1, 2001, pp. 4-7 [2] D. Zhang et al., ”OSGi based service infrastructure for context aware automotive telematics”, IEEE 59th Vehicular Technology Conference, May 17-19, 2004, pp. 2957-2961. [3] J.Ye et al., ”Representing and Manipulating Situation Hierarchies using Situation Lattices”, Revue d’Intelligence Articielle, vol. 10, no. 10, 2008, pp. 1-15. [4] C. Anagnostopoulos and S. Hadjiefthymiades, ”Enhancing situationaware systems through imprecise reasoning” IEEE Transactions on Mobile Computing, vol. 7, pp. 1153-1166, 2008. [5] N. Wei et al., ”An Ontology-Based Approach to Personalized SituationAware Mobile Service Supply”, Geoinformatica, vol. 10, no. 1, 2006, pp. 55-90. [6] S. Dobson, J. Ye, ”Using fibrations for situation identification”, Pervasive 2006 workshop proceedings, Springer Verlag, 2006, pp. 645-651. [7] A. W. Moore, ”Bayes Nets for representing and reasoning about uncertainty”, technical report, tutorial slides, 2001, Available: http: //www.cs.cmu.edu/∼awm/ [8] D. Heckerman, ”A tutorial on learning with Bayesian networks”, Technical Report MSR-TR-95-06, Microsoft Research, Redmond, June, 1996. [9] R. J. A. Little and D. B. Rubin, ”Statistical Analysis with Missing Data”, John Wiley & Sons, 2002. [10] M. J. Beal, ”Variational Algorithms for Approximate Bayesian Inference”, a dissertation for Ph. D degree, chapter 2, 2003. Available: http://www.cse.buffalo.edu/faculty/mbeal/thesis/index.html [11] H. Zhang, ”The optimality of Naive Bayes” in Proceedings of the Seventeenth FLAIRS, May 17-19, United States, 2004, pp. 562-567. [12] ”Google N-Gram-Patterns”. Available: http://n-gram-patterns. sourceforge.net/ [13] ”Drools is a business rule management system (BRMS) and an enhanced Rules Engine implementation”. Available: http://jboss.org/drools/
