Information Dissemination Between Mobile

Information Dissemination between Mobile Nodes for Collaborative Context Awareness
Christos Anagnostopoulos, Member, IEEE Computer Society, Stathes Hadjiefthymiades, Member, IEEE Computer Society, and Evangelos Zervas, Member, IEEE Computer Society
AbstractIn our everyday life, we frequently experience cases where persons group together. In those cases, context-aware systems capture and process identical context. Therefore, the need to collaboratively address Context Awareness (CA) emerges. In the considered setting, ad hoc networking between mobile nodes enables the exchange of information, thus, CA is facilitated. The synergy between mobile nodes materializes the Collaborative CA (CCA) paradigm. We advance the general CCA concept by performing communication between nodes probabilistically in a way similar to virus (epidemic) spreading. Nodes feature a hierarchical information model, which can be exploited by an information diffusion process. Multiple pieces of information, exchanged as epidemics, can complete the information present at a certain node, which in turn infers and spreads new information. We study this novel scheme extensively through an information model for context and an analytical framework (Markov process) with simulations. Our findings show that the information spreading large benefits from the mobility of nodes and semantic processing of the information model. Index TermsMobile computing, mobile applications, pervasive computing.
1 INTRODUCTION
E focus
on Context-Awareness (CA), which is the
ability of mobile applications to sense, interpret, and react to changes in the environment they are situated in [1]. In this paper, we deal with a mobile and distributed computing setting with the following characteristics: . A network of mobile nodes (satellites) and stationary nodes (hubs). . All nodes receive and relay contextual information (context) to other nodes. . Satellites attach to hubs stochastically. . Information is floating throughout the network in a probabilistic way. . Exchanged information has a certain quality and semantic value and presents dependencies to other pieces of information. . Semantic dependencies can be further exploited by nodes to more accurately tune their behavior, thus, forming the concept of Collaborative CA (CCA). We propose 1) a context model and 2) a context dissemination scheme for CCA. The context model is ultimately based on semantic dependencies among pieces of information that are distributed across the considered network. The information dissemination scheme is probabilistic, so that each node decides with a given probability whether to transmit/relay information or not. We capitalize on the similarity between information dissemination and the epidemic model, mostly in the abstractions of virus spreading, virus severity, virus interdependencies, and possible transmutations. The great advantage of epidemic-inspired communication is that dissemination proceeds on a local basis, without
any central coordination [2]. Such communication is highly reliable, efficient, and resilient to communication problems. However, the study on dissemination of dependent pieces of information is a major challenge. For instance, consider the case where persons group together, e.g., in conferences. Such persons share, at least temporarily, common interests and preferences, e.g., a group of persons interested in the same scientific presentation. Group members experience similar situations and many of their mobile computing devices (nodes) sense and process different pieces of information (e.g., a noise sensor or a localization device). The described setting, which is not so rare, introduces the need for treating CA in a collaborative manner. Collaboration denotes the synergy between neighboring nodes for sharing dependent pieces information. Such pieces of information can be further processed by receiving nodes to infer information of better semantic value. We adopt the epidemic-based information dissemination model. Information dissemination is handled by a simple, message passing protocol which operates under certain rules, e.g., spatio-temporal constraints. A piece of information should reach potential recipients within a certain time horizon. Moreover, such piece may be considered valid only in the neighborhood of the node (in specific geographical area or within a certain number of hops). This means that, not all information needs to be propagated everywhere. Through this paper, we show that the stochastic
1710 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 10, NO. 12, DECEMBER 2011
. C. Anagnostopoulos and S. Hadjiefthymiades are with the Department of Informatics and Telecommunications, National and Kapodistrian University of Athens, Panepostimioupolis, Ilissia 15784, Athens, Greece. E-mail: {bleu, shadj}@di.uoa.gr. . E. Zervas is with the Department of Electronics, Technological Educational Institute of Athens, Egaleo 12210, Athens, Greece. E-mail: zervas@ee.teiath.gr. Manuscript received 10 June 2009; revised 12 Oct. 2010; accepted 15 Dec. 2010; published online 2 Feb. 2011. For information on obtaining reprints of this article, please send e-mail to: tmc@computer.org, and reference IEEECS Log Number TMC-2009-06-0220. Digital Object Identifier no. 10.1109/TMC.2011.19.
1536-1233/11/$26.00 _ 2011 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS
nature of epidemic spreading achieves very high coverage, even, for low forwarding probability. The similarity of our model to epidemics extends beyond the spreading scheme as shown below. The basic scheme. A disseminated piece of information abstracts an epidemic in the sense that, a node carrying a piece of information becomes infected; otherwise it is susceptible. An infected node can disseminate information to its neighbors in a probabilistic way. When a node carries information, which has become invalid, then it becomes susceptible and can be reinfected at a later stage. The extended scheme. The severity of viruses is included in our abstraction. We introduce the metaphor of a stronger epidemic than the one, which has previously infected a given node. Stronger epidemics abstract more valuable information, i.e., information is valued more by the receiving node. To be able to assess the value of the incoming piece of information, each node provides a mechanism (see Section 4) which unambiguously specifies semantic interdependencies among pieces of information and evaluates them. We can figure out the following scenarios: . the incoming piece of information, denoted as p, is of better quality than the one previously received (e.g., more up-to-date message), and . the p can act as a seed, thus, enabling the receiving
end to generate/infer additional information. The p covers the missing elements needed for generating extra information, e.g., the combination of location, time, and information about the scheduled lectures of a conference could be exploited by a personal conference assistant application. We adopt the term information completion for this case, where some vital pieces of information are missing (until reception of p) in order to generate new information of higher semantic value. Since, in any of these cases, the derived information (i.e., after the reception of new information) is deemed more valuable for the receiving node, we abstract them as epidemic aggravation. The last part of our abstraction is related to dissemination of derived/inferred information. Imagine that the derived information is not kept locally in the node, but further treated as a virus and propagated to the neighborhood as new information. A new virus starts spreading and will, possibly, infect nodes in the proximity. Hence, in a group of mobile nodes multiple pieces of information circulate, thus, establishing the need of studying a multiepidemic information dissemination scheme. The contribution of this paper is twofold: the definition of a contextual information model and the study of the multiepidemic information dissemination. Our objective is to develop a hierarchical context model, which includes a (reasoning) mechanism for evaluating semantic dependencies among pieces of information. In addition, we develop a mathematical model that deals with the multiepidemic information dissemination in light of 1) information aggravation and 2) the mobility behavior of nodes. We deal with mobile and stationary nodes. The stationary nodes (hubs) form a social network. On the other hand, the mobile nodes (satellites) move between stationary nodes stochastically forming a hub-satellite like network. To this purpose: . We first abstract the problem of multiepidemics based on the fact that nodes reason about whether to replace existing pieces of information with more valuable incoming information or not. . We use a graph-based representation in order to formulate multiepidemic propagation through a spatio-temporal random process based on interactions between nodes. . We make use of a discrete-time model and network topology information to describe the spatial and temporal statistical dependencies among diverse epidemics. The temporal dependency is modeled as a Markov process. Moreover, we generalize the Markov model in [3], which deals with monoepidemic spreading.1 . We apply our models to describe the final status of infection, which corresponds to the equilibrium solution and characterizes the potential distribution of the epidemics over the network. Finally, we show that the multiepidemic spreading accelerates as long as satellites tend to move to diverse hubs with the intention of inferring and sharing context. The paper is organized as follows: Section 2 discusses the rationale of our abstraction and Section 3 reports prior work. Section 4 introduces the context model and Section 5 reports the analytical model for information dissemination. In Section 6, we report performance and comparative
assessment with certain models found in the literature, while Section 7 concludes the paper.
2 RATIONALE
According to the Kermack and McKendrick epidemic model, an individual can be in three states: Infected, Susceptible, (i.e., an individual is prone to be infected) and Removed (i.e., an individual is immune, as it recovered from the disease). This kind of model is usually referred to as Susceptible-Infected-Recovered (SIR). A simplified version of that model is the Susceptible-Infected-Susceptible (SIS) model in which an individual can only be found in the susceptible and infected states. This means that an individual never gets immune after its contact with the epidemic. Moreover, SIS assumes that an infected individual cannot be reinfected by another (transmuted) epidemic. In addition, in the Susceptible-Infected (SI) model, an individual never turns susceptible if infected once. The SI, SIR, and SIS models cannot abstract the phenomenon in which an individual 1) might return to the susceptible state from the infected state, 2) might be infected by a transmuted epidemic, and 3) never turns immune from an epidemic. By taking into account the semantic dependencies that contextual information typically exhibits, we move beyond a simple flat model (i.e., the SI, SIS, SIR epidemic models) to a layered information organization, which is also reflected to the virus abstractions adopted. We extend SIS at the point that an infected node with epidemic p can
ANAGNOSTOPOULOS ET AL.: INFORMATION DISSEMINATION BETWEEN MOBILE NODES FOR COLLABORATIVE CONTEXT AWARENESS 1711
1. The spreading of an independent piece of information.
be reinfected with a transmuted epidemic q resulting to the aggravation of the nodes condition. The q indicates a more infectious and contagious (stronger) virus than p. Then, the node may infect neighbors with more detailed information or information of better quality. On the other hand, the cure of a node refers to the improvement of its condition. The abstraction of cure indicates that, for instance, information is no longer valid (i.e., beyond the spatial scope or outofdate). We adopt the acronym SIaS for the extended SIS model indicating that the proposed model deals with reinfections of different epidemics; a stands for aggravation. Once no aggravation occurs, we obtain the SIS model. In SIaS there are numerous dependent pieces of information (epidemics) spread in the network simultaneously. Information of a high-quality level refers to a strong (transmuted) virus. The strongest epidemic has the potential to infect a large percentage of nodes, contrary to weaker epidemics, which infect a smaller portion of the group. The dependencies among epidemics are exploited by the nodes to substitute, complete, or generate new information. If a node has the ability to infer information then it can disseminate it into the group, thus, augmenting the current common knowledge. This reciprocity guarantees a modus operandi for a network of autonomous nodes working under the SaIS model for collaboratively sharing and inferring context. Consider the set of pieces of information H. Let us define a relation _over H such that p _ q denotes that p is of better quality than q. Quality of a piece of information can be, for instance, more detailed, more fresh, and more specific referring to the perspective of interest/interpretation of the node. Each node compares incoming p 2 H with existing q 2 H. If p _ q then the node accepts the incoming p and its condition is aggravated, otherwise, the node does
not replace q with p. Each node can, further, locally refine information independently of other nodes. It is also possible for a node to exploit the received p in order to generate more valuable information _ (aggravation from p to _) and disseminate it across the network. Example. We provide an example that maps the multiepidemic abstraction to real contextual information dissemination among location-aware applications. Throughout the paper, we regularly refer to this example to further explain the mapping of the concepts to the SaIS model. At first, in this example, the basic context type is position information, so the p; q epidemics refer to location information and the relation p _ q indicates that location p is more accurate and detailed than q. Second, we make use of a short range communication scheme between mobile nodes through wireless interface (Wi-Fi network). We also consider visitors in a museum who are equipped with smart-phones. They are visiting an exhibition held in the first two floors of the museum. Not all the smartphones of visitors are equipped with the same localization technology (e.g., terrestrial localization like EOTD, Infraredbeacon, WLAN positioning) or have Wi-Fi network connectivity. The position information is exploited by a CA application in order to present to the end-user (visitor) contextual information, e.g., a virtual museum tour. A visitor estimates her position through the EOTD localization method ([8]). Hence, her context is p0 contextlocation is 0Museum building :EOTD0. That node disseminates p0 to her neighborhood (other visitors) denoting that, her current location is determined with EOTD accuracy along with a time validity indicator (e.g., 2 minutes). At the current time, some visitors can receive context p0 together with temporal and spatial constraints. Another neighboring visitor estimates with a higher accuracy, her position through the Wi-Fi received signal strength (RSS), thus, the information p1 contextlocation is 0Museum building; Floor 2:WLAN0. p1 is also disseminated to neighboring nodes. The visitors on Floor 2 that have been recently infected by p0 can now be reinfected with p1 since the latter information describes the location of nodes in more detail. In addition, a node infected by p1 can also infer p0. Earlier, another neighboring visitor exploits the incoming p1 and estimates a more accurate position, e.g., p2 contextlocation is 0Museum building; Floor 2; Room 3: Infrared0, since an Infrared receiver is mounted on that node [9]. The latter localization estimation (p2) is far more accurate than p1 and p0. The node disseminates p2 as an improvement on p1 to her immediate local neighbors, for example, a certain number of visitors (inside Room 3). The neighboring visitors of that node can now be supported with more detailed information, thus, improving the quality of the service seen by users. It holds true that p2 _ p1 _ p0. Another visitor, which has Wi-Fi internet access and has received p2, can now extend her knowledge about her surroundings: a lecture on a specific drawing is scheduled at 11:00 a.m. in Room 3. This visitor combines current location, time, and information about the prescheduled lectures of the museum (e.g., starting time, presenters profile) and infers that the lecture on that specific drawing will be held in that specific room in the next 15 minutes ( p3 contextactivity is 0lecture on a drawing in Floor 2; Room 3; at 11:00 a:m:; presented by a well-known artist . . . 0) . The inferred information p3 (a refinement of p2) refers to an
expansion of the up to now local knowledge of the visitors (epidemic aggravation). The visitor disseminates p3 within the group of interested visitors in an area close to Room 3 (or of the second floor). This results to the improvement of the understanding of the surroundings for the interested visitors. Moreover, a visitor, which is located in Room 4 of the second floor heads for Room 2 passing through Room 3 (in order to meet a friend). This visitor can disseminate p3 (or an abstract description of p3, e.g., lecture on Room 3) for a certain time to possibly interested visitors located in Room 2 as long as the visitor receives p3 from visitors of Room 3. p3 can be also disseminated to other rooms or floors because of visitors mobility. Once visitors move among rooms, they can disseminate pieces of information of different refinement levels. Finally, it holds true that p3 _ p2 _ p1 _ p0.
3 RELATED W ORK
Several analytical models have been proposed in monoepidemic information dissemination in static networks. Such models focus mainly on certain theoretical aspects of epidemic-based information dissemination (i.e., epidemic
threshold, impact of static network topology on the dissemination process), especially, for the SIS model. Specifically, the work in [4] presents an analytical epidemicbased information dissemination model for homogeneous network topologies. The key result from this model is the prediction of the monoepidemic spreading and the estimation of the monoepidemic threshold. In more complex network topologies like scale-free networks, the authors in [21] studied the monoepidemic spreading. The key outcome of this study is that the epidemic threshold 2 in such network types does not exist. On the other hand, the authors in [22] proposed a SIS information dissemination model over arbitrary network topologies. The major result of this study is the estimation of the epidemic threshold in information dissemination w.r.t. any static network topology by adopting an eigenvalue approach. The epidemic threshold in this case depends on the infection and cure rate. It is worth noting that the discussed analytical models deal with the information dissemination process in networks of stationary nodes. However, such models do not capture the idea of multiepidemic spreading and are not appropriate for modeling epidemics in nonhomogeneous mobile networks. Therefore, the research on the nonhomogeneous networks [25] is not applicable to mobile networks for the reasons discussed in [17]. The epidemic frameworks for power-law networks do not cover the concept of node mobility. Moreover, given the adjacency matrix A of an arbitrary network topology, the epidemic threshold depends on the largest eigenvalue of A as shown in [22]. In mobile environments, A and the corresponding eigenvalues change over time, thus, it is unclear how to construct A with time-dependent values. In addition, node mobility introduces nonhomogeneous connectivity distributions (as in our satellite-hub network), thus, they cannot be represented by a simple average of node connectivity. Mobility also introduces continuous change in the neighborhood of each node. We mainly base our analytical information dissemination model in [3], which we extend in order to study collaborative context-awareness in mobile environments. Specifically, in [3], the authors analyze a Markov processbased
framework that characterizes the spreading of epidemics through the SIS model. In this model, the authors deal with the prediction of the information dissemination using the SIS model, in which the resulting analytical stochastic model matches well with the simulation results. We adopt the stochastic method of the model in [3] in order to define and develop the multiepidemic-based information dissemination through a Markov process. As long as the model in [3] does not capture the concept of aggravation, we extend it by defining the reinfection to different levels and introduce the corresponding analytical equations for the Markov process. In addition, the model in [3] does not consider node mobility, which is a very crucial component of our work. Besides the analytical information dissemination models, we have to consider the concept of reinfection in order to achieve more detailed collaborative context. The following related models and frameworks deal with the main issues of CCA and the corresponding key findings are reported. Specifically, the authors in [26] propose an approach for collaborative context dissemination among groups of nodes. Such architecture tunes the dissemination process with the reliability of information. The reliability degree of a piece of information is an important indicator. Information is accepted by a node when the reliability indicator is high. Our model can represent such indicator through the semantic relationship between pieces of data. However, the model in [26] does not consider reinfection, thus, more detailed context cannot be inferred. In addition, the dissemination scheme discussed in [26] is a simple floodbased model. The performance handicap of a flood-based approach w.r.t. epidemic spreading is revealed in our performance assessment (Section 6.3) and in [23]. Finally, the model in [26] does not assess the dissemination process w.r.t. node mobility. In addition, the work in [24] deals with the application of the SI model in mobile networks. Nonetheless, as discussed above, the SI model cannot be considered appropriate for collaborative context dissemination, since infected nodes cannot recover or get reinfected with more detailed context. Moreover, the autonomous gossiping algorithm in [6] refers to the selectivity attribute of the epidemic spreading and the model in [7] refers only to spatially constrained information dissemination. The key finding in the latter model is that information spreads nearby nodes with higher probability than to distant nodes. However, such algorithms do not consider context dissemination, thus, nodes cannot generate new information and, consequently, cannot augment knowledge. In addition, the impact of the mobility of nodes is not studied in [6] and [7]. The work in [17] describes a monoepidemic information dissemination model, which uses exact location information for treating mobility of nodes. The major difference between our model and that in [17] is that our model interprets the concept of mobility as the change of the connectivity between nodes and not as detailed location information, as dealt in [17] (longitude and latitude coordinates). In a mobile environment of CA applications, it is not always possible for an application to be supported with exact location information (consider our indoor network example). Moreover, the model in [17] does not support the refinement of contextual information in context dissemination. In [30], the authors study the control of the epidemic dissemination in mobile, ad hoc networks. The key result of
the proposed analytical model is the control of the reliability level on message passing based on the degree of connectivity of hosts (assuming homogeneous networks). This model deals only with monoepidemic spreading focusing on the precise tuning of the infection rate. Such functionality can be adopted by our model in controlling only the dissemination of the most detailed context, thus, optimizing the efficiency of information dissemination. Moreover, a simple concrete data reasoning with data-filtering operators have been proposed in the policy-driven context dissemination model discussed in [31]. Such model deals with overflow conditions in context dissemination based on set/comparison filters. The key finding from this study is that the adoption of such filters enables the receiving node to either accept or reject incoming context. However, such model lacks semantic relations (e.g., generalization relation , is-a taxonomy), thus,
2. Epidemic threshold is an indicator that denotes the ability of an epidemic to be pandemic or to die off in a network of individuals.
the reasoning about the acceptance of hierarchically dependent pieces of context cannot be supported. The CoCA framework in [32] deals with a context management mechanism that adopts the Resource Description Framework (RDF). However, several conceptual models and more complex schemes that support context inference and reasoning have been proposed in pervasive environments [12]. Especially, in [28] the authors exploit enhanced context semantics (OWL-DL) than the adopted RDF in [32] when dealing with adaptivity for pervasive environments. In addition, our proposed context model can support any hierarchical scheme (e.g., is-a taxonomy). Based on the taxonomy of the collaborative context-aware systems in [33], we can categorize the characteristics of our model into the three proposed aspects in [33]; goal: the nodes aim at achieving a common/similar depiction of the environment, approach: consensus-free, i.e., the nodes take local decisions on the current context sharing diverse levels of context, means: nodes sense, fuse, infer, reason, and share current estimations of context. In addition, the authors in [29] proposed a model for collaborative exchange of context in random network topologies. In such model [29], the nodes may not always have the capability to sense their environment and infer their exact context. Thus, the applications carried by such nodes may not be able to obtain the needed input for execution. Finally, in our preliminary work [16] we propose an analytical contextual information dissemination model based on a double-epidemic SIS model (i.e., dissemination of two levels of infection, K 2) in static homogeneous networks. Specifically, in this model we deal with dissemination of multimedia semantics. The disseminated information covers software/multimedia content metadata, semantic content specification files (e.g., Composite Capabilities/ Preference Profiles (CC/PP)) and mpeg7 files. Based on such model, nodes exploit multimedia semantics and, thus, share semantically enriched multimedia content among neighboring nodes. The proposed context model in [16] uses ontological representation. A node receives the desired multimedia content that matches its interests once the corresponding semantic description (epidemic) propagates across the network. In this paper, we extend the model presented in [16] as follows: 1. we study the dissemination of multiple, semantically
dependent epidemics (K >2) and, for each infection level, we generalize the analytical equations w.r.t. epidemic prevalence and decay, 2. we study the behavior of the multiepidemic dissemination concerning static (small-world network) and mobile network topologies (satellite-hub network), 3. we study and compare the efficiency (see Section 6) of the multiepidemic model against a range of other relevant schemes, and 4. we extend the reasoner process [16] for K >2 epidemic levels adopting a hierarchical scheme for contextual information (i.e., generalization relation).
4 CONTEXTUAL INFORMATION MODEL

A well-known definition of context is the following: context is any information that can be used to characterize the situation of an entity. An entity is a person, place, or object that is considered relevant to the integration between a user and an application, including the user and the application themselves [5]. Context refers to the current values of specific parameters that represent a nodes activity (e.g., moving in a building), event (attending a lecture), and environmental information (e.g., illumination) in a specific place and time. Inferred context is the additional information that can be deduced from the current context. Nodes incapable of inferring (new) context can possibly acquire such knowledge from neighbors. The terms context and information are used interchangeably according to the meaning.
4.1 Context Representation Context representation is complicated due to the fact that different nodes value context differently. Different types of context might appear, e.g., the context of each node in a group (individual context), the context of the group itself (collaborative context), and the context of the task [10], [11]. The context determined by a sender node might be differently interpreted by the receiving node. Hence, the collaborative context _ does not, necessarily, imply that all nodes of that group share individually identical context. Instead, the context p of a node can be at least as valuable as _ (p _ _). Nevertheless, each node can further locally refine context independently of the other nodes, called as individual context. The semantic structure of the proposed information model defines the criteria of distinguishing the collaborative context from the individual context. The hierarchical knowledge representation is adopted. The interpretation of the dependency _ _ p is application specific, for instance, 1) _ is of better quality than p, and 2) _ represents more detailed context than p, thus, one deduces p from _. Further discussion on context modeling and representation can be found in [12] and [13]. Let Pn be the finite set, Pn fp1n; . . . ; pkng, of k parameters pin; i 1; . . . ; k of level n _ 0, which assume values oi in the domain Dompi . There can be different parameters in different levels. A parameter pin is instantiated at time t if, at time t, a value oi 2 Dompi is assigned to pin, that is pin oi. Context of n-refinement level is the set pn fo1; . . . ; okg of instantiated parameters. The parameters belonging to P0 (0-level) represent basic context; context that cannot be inferred by other parameters belonging to P0 (e.g., sensor readings, p10 position; p20 time; p30 illumination; P0 fp10; p20; p30g). Consider pn and pm; n 6 m. Then, context completion of pn with pm is the expansion of pn with
distinct values of pm resulting to the formation of context of n_-refinement level, that is pn_ pn [ pm n pn; n_ maxn;m: As long as pn_ is completed with additional values, the CCA application might be able to infer additional context. Consider the N-ary logical synthesis f _ Pn1__ _ _ _ PnN. Then, at time t, inferred context pm 2 Pm of m-refinement level, m > 0, is the result of the conjunctive synthesis f on pini 2 Pni; i 1; . . .;N; ni < m: f p1n1; . . . ; pNnN is represented as the antecedent-part of the implication ! and inferred context pm as the
consequent-part. The antecedent-part refers to the proposition pini oi or simply pi oi, oi 2 Dompi . The implication ! concludes pm of a higher level set with m maxni 1; i 1; . . .;N, that is p1 o1 ^ _ _ _ ^ pN oN ! pm: 1 The pm is the classification of the pattern fp1 o1; . . . ; pN oNg, where the elements pi oi can be formed by context completion. Evidently, the higher the value of m is the more detailed information is obtained. Details on a more complex inference can be found in [28] and [14].
4.2 Context Reasoning Consider the inferred context pn 2 Pn and qm 2 Pm of levels n < m, and let Apn and Aqm, and op i2 Dompi and oq i 2 Domqi , be the set of antecedents and the values of pn and qm, respectively. Then, inferred context can be hierarchically structured forming specialization relations _ with interpretation as follows: Apn _ Aqm ^ _ op i oq
i
_ $ qm _ pn; i 1; . . . ; jApnj. p pn represents more abstract context than q qm but the opposite implication does not always hold true. If _p is the set of all pieces of context that are more abstract than p, i.e., _p fejp _ e _ e qg, then a context taxonomy H is H fq _ pj_p \ _q 6 ;__p _ _qg: A node i can autonomously reason about whether to accept incoming context sent by a neighbor or not based on the reasoner algorithm in Table 1. Specifically, consider the hierarchy H f_ _ q; q _ p; p _ g and the fact that node i receives p at time t. Then, node i: 1) is likely to be reinfected with q, such that, q _ p at time t 1 (aggravation from p to q) or 2) can infer p from q. In both cases, node i accepts q and, thus, the condition of node i is aggravated. If the nonoccurrence of p holds true at time t, that is, :p ! true, then the nonoccurrence of q is also concluded, i.e., :q ! true.3 This leads to the fact that, a node partially recovers from any context _ 2 H as long as _ _ q _ p and remains infected by any context 2 H as long as q _ p _ . In this case, the condition of node i is improved. Finally, if __ ;, node i fully recovers. During the (inferred) context dissemination, it is assumed that all the disseminated pieces of context are temporarily valid and constrained to a certain number of hops (or to a certain geographical area). The application of SaIS focuses on how context is disseminated based on context completion,
the inference capability, and the mobility of nodes. However, timing issues related to the temporal validity of the disseminated context can be handled through SaIS. In that case, there might be two types of the relation _ _ p: a temporal relation denoting that _ is measured more recently than p, but this does not necessarily imply that p is obsolete (evidently, an obsolete piece of context is not further propagated), and a refinement relation denoting that _ is more detailed than p. A node, which obtains p, may accept _ or not based on the decision of what is preferable: up-to-date context at the expense of more detailed context than the existing, or any combination of these two criteria. Finally, a group of nodes is considered infected when each node receives the most detailed context of the groupcollaborative context _, i.e., there is no context # _ _. The basic interpretation of the cure rate is that it denotes the rate at which context p turns invalid on a node under certain constraints (e.g., temporal and spatial constraints). In other words, the cure rate is the reciprocal of the time validity. In our example, the cure rate _lk; l > k _ 0 denotes that if position information pl is not valid (for a certain time horizon) then the nodes position information refers to position pk that holds true and pl _ pk (see Lines 6-11 in Table 1). This means that, the nodes position is pk, which is valid but less detailed than pl. Once, pk position information turns invalid for the node then, automatically, pl is also invalid. However, the opposite case is not always true, that is, if pl is not valid, then, pk might be valid independently of pl. That is because we can conclude pl from pk and vice versa. In case k 0, then no context is valid for the node, thus, each detailed/inferred context pl is certainly invalid. Back to our example, p3 context is valid once, at least, p2 is valid. The corresponding cure rate _32 denotes the rate at which p3 turns invalid while p2 is valid. We also provide another example derived from our previous work in [16]. In this work, we deal with the double-epidemic model (i.e., K 2), in which the contextual information refers to user preference on multimedia content and multimedia content profile. The semantic relation p _ q denotes that the user preference q matches with a multimedia profile p. In this case, the cure rates in [16] denote the rate at which a profile matching between user preference and multimedia content profile is 1) impossible between levels 0-1 and 1-2 (p1 _ p0 or p2 _ p1 with _ _21 _10) or 2) definitely impossible (p2 _ p0 with _20 _). That is, if the user reduces his/her preferences then he/she obtains multimedia content of lower capability. On the other hand, if he/she increases his/her preferences then he/she might obtain no multimedia content (because of no matching).
5 INFORMATION DISSEMINATION MODEL

5.1 Notation We use a directed graph GV ;E to represent a network with multiepidemic propagation, where V is the set of nodes and E is the set of edges. A neighborhood of node m, denoted by Vm, is a subset of V where every node i in this
TABLE 1 Algorithm: Node Reasoner

3. If, however, both p and q can be either true or false (bivalent) and p can only be true if q is true then modus tollens stands, i.e., q ! p ^:p ! :q.
subset has an edge connecting to node m, i.e., Vm fi j i;m 2 Eg. Each edge i;m 2 E is associated with _im, which is the infection rate (forwarding probability)
with which an infected node i can infect (relay information to) a neighboring node m. The state of node i at time instant t assumes K 1 values which are represented by the K 1-dimensional vectors pk 0; . . . ; 0; 1; 0; . . . 0_T where all values are zero except the kth component (k 0; 1; . . .;K) which takes the value 1. A state of value pk denotes that the node is in infected state of level k. The most detailed information that is disseminated across a network refers to inferred context of refinement level K. A node is either in susceptible state p0 or in some infected state with epidemic (context) pk. We deal with the expected percentage nkt of infected nodes for all epidemics pk, k 1; . . .;K. A node i of infection level pl can infect a neighboring node j, which is in state pk, iff pl _ pk, i.e., either node j is susceptible (p0) or infected at a level lower than l. A node i of infection level pk can be cured in two ways: 1) in the full cure case, node i transits in one step to susceptible state p0 with rate _k0; k 1; . . .;K (see Fig. 1); 2) in the partial cure case, node i transits in a lower infectious state pl, 1 l < k, with rate _kl. In partial cure, the node may be still infected since it transits to a state of weaker epidemic than the previous one; a partially cured node can be also reinfected with a stronger epidemic. Fig. 1 shows the state transition diagram of a node. It should be noted that the transition rates Qkl depend on the number of neighboring infected nodes, their infection level, and the infection rate _ij; the Qkl values will be further clarified in the following paragraphs.
5.2 Assumptions Mobile networks vary widely in their characteristics due to the mobility of nodes. Inhomogeneous connectivity distributions are often encountered in such environments [17]. Such distributions range from nomadic networks, in which nodes relocate forming interconnected clusters of nodes, to ad hoc networks, in which nodes move freely and fail to form some form of structure. Some nodes may have more communication links (attract more nodes) than others. In our analysis, there are nodes that capture and infer more information w.r.t. semantic value than other nodes, thus, being more popular than others. Such nodes, called hubs, generate nonuniform connectivity distributions. That is because, a node, called satellite, which attaches to a very popular hub, will have many more neighbors in transmission range than a satellite attached to a less popular hub, as illustrated in Fig. 2. This mobile network can be modeled by multiple clusters of hubs and satellites (attached to hubs), which yields the nonuniform connectivity typically encountered in mobile environments. Each cluster is represented as an island and, as satellites roam between islands, they share information with them. We adopt a small-world (social-network) network topology among hubs. Hubs may infect each other. We assume that the satellites are attached only to one hub at a time and infect (or be infected by) that hub. Nodes are equipped with short range communication mechanisms. Neighboring satellites communicate through their common hub, thus, forming a star-like scheme. A satellite influences the visited island by stochastically infecting the corresponding hub and the attached satellites. Moreover, a hub is infected not only by the attracted satellites but also by neighboring hubs. Due to the mobility of satellites In such topology, the average connectivity of a node, represented by hwi, is different among hubs and
satellites. Due to the mobility of satellites, a hub may have a number of satellites ranging from 0 to hwi. Therefore, SaIS does not strictly rely on the average connectivity statistics, as supported by the most used epidemic models [4]. SaIS models node mobility through the random changes in connectivity among nodes. However, in [15] it is reported that some inhomogeneous networks have connectivities guided by power-law topologies, where the probability of a node i to have wi degree of connectivity is proportional to w_ i for some _ 2 2; 3_. If there are hubs that are connected with a few nodes into an island, connectivities may resemble to those of a power-law network. However, the epidemic framework for power-law networks does not cover node mobility, thus, such degree distributions cannot be adopted [17]. In this work, we examine the mobility-dependent differences in effective virulence of multiepidemic dissemination. Satellites with high mobility are better mixed than satellites with low mobilityduring a given period of time. Satellites communicate with a wider set of neighbors than hubs. Higher mixing rates boost multiepidemic spreading since infected nodes have more opportunities to communicate with susceptible nodes. We assume M hubs hm, m 1; . . .;M that form a network with adjacency matrix A. Satellites are attached
Fig. 1. State transition diagram of a node with partial cure rates _kl;K _ k > l > 0, and full cure rates _k0;K k (not all possible transitions are illustrated for readability reasons). Fig. 2. Satellites roam between islands changing the network topology.
to each hub hm. The number of satellites attached to hub hm is time dependent and denoted by Smt. These satellites are denoted by Si;m, where the subscript m denotes the hub and i is the index of the satellite, i 0; . . . ; Smt. The connectivity among neighboring hubs remains unaffected. On the other hand, satellites change their communication connectivity among hubs. A satellite possibly infects only the hub it is attached to, with infection rate _sh (which is the same for all satellites). On the other hand, a hub may infect either its neighboring hubs with infection rate _hh or the attached satellites with infection rate _hs. The total number of satellites in the network is S PM m1 Smt. A satellite Si;m remains connected to hub hm for an exponentially distributed time Ti;m _ expf_0g with parameter _0. Let T be a random variable describing the elapsed time before a network topology change occurs due to a satellite movement, that is T mini;mTi;m. Then, T is exponentially distributed with parameter _ _0S. Hence, we have that T _ expf_0PM m1 Smtg. Furthermore, we can assume that the system evolves in small time intervals of length d. That is, possible infections take place at the end of the time intervals of length d. To this end, the probability that time T takes value in the interval nd; n 1d_ is Pfnd < T n 1dg 1 e_de_dn; 2 which, as expected, is the geometric distribution with success probability 1 q 1 e_d. From now on, we normalize time by setting d 1. Hence, the probability that a move takes place in an interval of length d, referred to as mobility probability, is q e_d: 3
If a satellite moves, then the originating hub ho is selected with probability Pfho mog Smo t PM m1 Smt ; 4 which implies that all satellites are uniformly selected as candidates for a move. The destination hub hd is selected according to Pfhd mdjho mog fhmd; Smd t: 5 The f_ function denotes how probable the hub hmd is as being the destination hub to which Smd t satellites are attached at time t. If a homogeneous distribution of satellites to hubs is assumed then each hub is equiprobable, that is, fhmd; Smd t M1. If satellites are gathered to the most popular hubs, i.e., to those hubs that obtain more detailed context, then (under a nonhomogeneous network topology) fhmd; Smd t w P hmd M m1 whm ; where whmd is the degree of connectivity of hub hmd. It is clear that, the pair of the originating and destination hub is selected with probability Pfho; hd mo;mdg Smo t PM m1 Smt fhmd; Smd t: 6
5.3 Multiepidemic Information Dissemination Model Let xi;mt and ymt denote the state of the satellite Si;m and of the hub hm, respectively, at time t. Once a node i (satellite or hub) can be infected only by its neighbors, the state xi;mt (or ymt) is statistically dependent on the status of its neighbors and xi;mt 1 (or ymt 1). Since the status of a neighbor also depends on its own neighbors then the status of all nodes is statistically dependent in space and time. Let the vectors xt and yt denote the status of the satellites and hubs at time t, that is xtxi;mt_>; ytymt_>; i 1; . . . ; Smt;m 1; . . .;M. It is clear that, such vectors are spatio-temporal processes. The nodes are categorized into satellites and hubs, thus, there are MS different configurations of the network; otherwise this number reduces to SM1
S
__ . Nevertheless, for each configuration the number of different states is K 1SM and, therefore, the total number of states of the Markov process is prohibitively large. To simplify the analysis, we first consider an homogeneous distribution for satellites. That is, for each hub m the distribution of the state xi;mt depends only on m, i.e., xi;mt xmt. This assumption is justified by the fact that in a quasi-stationary environment, that is, rare satellite moves compared to the number of infection intervals of length d, the hubs and, therefore, the attached satellites, reach their steady state probabilities in a few steps. A second consideration involves the mobility of the satellites. Suppose that a satellite leaves hub mo and moves to hub md, which has already Smd t satellites attached to it. Then, as a common distribution of the satellites attached to
hub md at time t, we take the weighted average Pfxmd t 1 plg 1 Smd t 1 Smd tPfxmd t plg Pfxmo t plg: 7 The basic objective of our analysis is the calculation of the probability Pfxmt pkg and Pfymt pkg, that is the probabilities that a satellite and a hub be infected with pk epidemic, respectively. Let Et denote the event that a satellite moves in the interval t 1; t_, and _ Et the complementary event. By using this notation, we have Pfxmt pkg q _ Pfxmt pk j _ Etg 1 q _ Pfxmt pk jEtg; Pfymt pkg q _ Pfymt pk j _ Etg 1 q _ Pfymt pk jEtg; 8 where q is the probability value calculated in (3). We then have to find the recursive formulas for Pfxmt pk j _ Etg, Pfymt pk j _ Etg and Pfxmt pk jEtg, Pfymt pk jEtg. For doing so, we treat the cases of nomove (or static) and move separately and drop the conditional events for notation simplicity.
5.3.1 The Static Case We analyze the probabilities of infections Pfxmt pk j _ Etg and Pfymt pk j _ Etg for satellites and hubs, respectively, assuming a static environment. The analytical model for the Static case is an extension of a certain part of the model in [3] (see Section 3). Specifically, we extend the probability
equations by considering multiepidemic information dissemination and conditional probability independence among nodes. Infection probability for a satellite. Given that no topology change occurs, i.e., no satellite moves, the probability that a satellite m is in state pk at time instant t 1, is Pfxmt 1 pkg X
pl
X
pj
Pfxmt 1 pk jymt pj; xmt plg _ Pfymt pj jxmt plg _ Pfxmt plg: 9 A common assumption that simplifies the analysis is that of conditional probability independence among nodes, that is Pfymt pj jxmt plg Pfymt pjg: 10 At first, let the node i have jVij neighbors. Then, based on the conditional probability independence, the total number of states needed to express the probability Pmt is reduced from OK 1jVi j to OjVij. The conditional probability independence in (10) explicitly involves detailed topology information and interactions among neighboring nodes. Second, we focus on the probability in (9) Ql;k;m Pfxmt 1 pk jymt pj; xmt plg; which expresses the probability of a satellite m to be infected from epidemic of level l to level k from its attached hub. Clearly, Ql;k;m _lk; l> k; 1
X
n<k
_kn ! 1 _hspTj
_
P
n>k
pn;
l k;
1 X
n<k
_kn ! 1 1 _hspTj _pk ; l<k; 8>>>>>>< >>>>>>: 11 where _hs is the rate of infection of a hub for its attached satellites. The following cases can be distinguished depending on the values of l and k: . Satellite m remains susceptible (Q0;0;m): in this case l k 0, which means that the satellite m is not infected by its attached hub. . Satellite m remains infected at the same level pk: that is, 0 < l k. Specifically, the satellite is neither infected with a stronger epidemic, nor turns susceptible (fully cure) or transits to an epidemic of lower level (partial cure). . Satellite m is infected with stronger epidemic pk: in this case 0 < l < k, which means that the satellite is infected with a stronger epidemic pk and is not partially or fully cured, . Satellite m remains in the highest level of infection: in this case l k K, denoting that m is the most infectious node, thus, QK;K;m 1 P n<K _Kn, . Satellite m is partially or fully cured: if k 0, then the satellite is fully cured; otherwise, the satellite is partially cured, i.e., transits to state k of lower level of infection than that of level l. In our example, we consider that a satellite m either 1) discards the position information, e.g., p2, once it is invalid, thus, corresponding to the cure rate _20, or 2) considers the position information of accuracy level k (e.g., p1), thus, corresponding to the cure rate _lk; k > 0. For instance, if p2 does not represent the position of node/visitor m at some time, then possibly its position can be either represented by p1 (cure rate _21) or p0 (cure rate _20). By combining (9), (10), (11) we obtain the recursive equation denoting the probability of a satellite m being infected at time t 1, that is Pfxmt 1 pkg X
pl
X
pj
Ql;k;mPfymt pjgPfxmt plg: 12 Infection probability for a hub. Now, we turn to the calculation of the probability of a hub being infected with epidemic pk at time t 1, that is Pfymt 1 pkg. Let the
neighborhood of hub hm at time t be denoted by Vmt. This neighborhood consists of satellites and adjacent hubs, i.e., Vmt Vm;ht [ Vm;st. Under the no topology change case, we have that Pfymt 1 pkg X
pl
X
yVmt
Pfymt 1 pk jYVmt yVmt; ymt plg; PfYVmt yVmt jymt plgPfymt plg; 13 where, the random vector YVmt denotes the status of all neighboring nodes of the hub m, that is, YVmt yit; xjt; i 2 Vm;h; j 2 Vm;s_; and yVmt yVm;h t; yVm;s t; is a realization of YVmt. Using the conditional probability independence assumption, we obtain PfYVmt yVmt jymt plg PfYVmt yVmtg PfYVm;h t yVm;h tgPfYVm;s t yVm;s tg PfYVm;h t yVm;h tg YSmt
j1
Pfxj;mtg: 14 Let us define the probability Rl;k that a hub m transits from epidemic pl to epidemic pk taking into account the current status of its neighbors (attached satellites and adjacent hubs), that is, Rl;k Pfymt 1 pk jYVmt yVmt; ymt plg: We can distinguish the following case depending on the values of l and k: . The hub is partially or fully cured, i.e., l > k. In this case, the hub m is partially (k > 0) or fully (k 0) cured, thus, Rl;k _lk; l > k: 15 . The hub remains infected at the same level, i.e., l k.
In this case, the hub m remains infected with epidemic pk and is not partially or fully cured. The interpretation of the partial and/or full cure case for a hub is the same as that for a satellite. Let _sh and _hh be the infection rates that a satellite infects its attached hub, and that a hub infects its neighboring hub, respectively. Then, we obtain that Rk;k 1 X
n<k
_kn ! _ Y
j2Vm;h
1 _hhyTj
t_
P
n>k
pn
_ YSmt
j1
1 _shxT
j;mt_
P
n>k
pn :
16 Equation (16) denotes that, a hub m remains infected at level k if 1) it is not fully of partially cured (the 1 P n<k _kn factor) and 2) is not infected by any other neighboring hub or attached satellite, which the neighboring hub or satellite iQs infected with a stronger epidemic than pk (factors
j2Vm;h
1 _hhyTj
t _
P
n>k
pn and
QSmt j1 1 _shxT
j;mt
P
n>k
pn ,
respectively). . The hub is infected with stronger epidemic, i.e., l<k In this case, the hub m is infected by a stronger epidemic pk than pl. Hence, Rl;k 1 Y
j2Vm;h
1_hhyTj
t_pk
YSmt
j1
1_shxT
j;mt_pk
0 @ 1 A_ Y
j2Vm;h
1_hhyTj
t_
P
n>k
pn _
YSmt
j1
1_shxT
j;mt_
P
n>k
pn :
17 Equation (17) indicates the probability of a hub to transit from infection level l to k. This means that the hub is only infected by the epidemic of level k and is not infected by any other epidemic pn with pn _ pk. By substituting (14-17) in (13), we obtain the recursive equation denoting the probability of a hub m being infected by epidemic pk at time t 1, that is, Pfymt 1 pkg X
pl
X
yVm;h t
X
yVm;s t
Rl;k _ PfYVm;h t yVm;h tg YSmt

j1
Pfxj;mtgPfymt plg: 18
5.3.2 The Dynamic Case We study the case in which the satellites move between hubs, thus, altering the network topology. A satellite, which moves between clusters of hubs, either relays context or is subject to aggravation. Since the satellites select the candidate hubs at random, there might be periods of time during which most hubs have very few attached satellites. The analysis of the dynamic case involves the calculation of the probabilities of infection with pk for satellites and hubs, i.e., Pfxmt pk jEtg and Pfymt pk jEtg, respectively, w.r.t. the mobility of satellites. We drop the conditional event of movement Et for notation simplicity. Infection probabilities for the originating and destination hubs. At first, we define the indicator function I_;m as I_;m 1; _ m; 0; otherwise: _ The I_; _ function will be used for indicating whether a satellite moves to a different hub or not. Actually, a satellite movement involves the originating hub ho and the destination hub hd. The probability of a hub infection with pk due to a satellite movement between ho and hd is Pfymt 1 pkg XM
mo1
XM
md1 md6mo
Pfmo;mdgPfymt 1 pk jmo;mdg; 19 where the event mo;md indicates that the satellite moves from the originating hub (mo) to the destination hub (md). By adopting the indicator function I_;m, we obtain Pfymt 1 pk jmo;mdg Pfymo t 1 pkgIm;mo Pfymd t 1 pkgIm;md Pfymt 1 pkg1 Im;mo Im;md: 20 The factor 1 Im;mo Im;md is either 0 (if m md or m m0) or 1 (if m 6 m0 and m 6 md). In addition, the probability Pfymt 1 pkg in the last line of (20) is calculated using (18). Clearly, for the originating hub, we have Pfymo t 1 pkg X
pl
X
yVmo;h t
X
yVmo;s t
Rl;k _ PfYVmo;h t yVmo;h tg _ SmYo t1

j1
Pfxj;mtgPfymo t plg: 21
Note that the neighborhood Vmo;st consists of Smo t 1 satellites, since a satellite leaves the hub ho. In the same way, for the destination hub we have Pfymd t 1 pkg X
pl
X
yVmd;h t
X
yVmd;s t[x;mo t
Rl;k _ PfYVmd;h t yVmd;h tg _ Y Smd t

j1
Pfxj;md tgPfx;mo tgPfymd t plg: 22 Infection probability for moving satellite. We now turn to the calculation of Pfxi;mt 1 pkg, i.e., the probability of infection related to a satellite i attached to hub m. We have that Pfxi;mt 1 pkg XM
mo1
XM
md1 md6mo
Pfmo;mdgPfxi;mt 1 pk jmo;mdg: 23
The cases m 6 md are treated using (12). These cases refer to the satellites that do not move to hub md. The only interesting case is that of m md, in which case we use (8). We update Pfxi;md t 1 pk jmo;mdg through the following formula: Pfxi;md t 1 pk jmo;mdg X
pl
X
pj
Ql;k;mdPfymd t pjg _ 1 Smd t 1 Smd tPfxi;md t plg Pfxi;mo t plg __ : 24 The (24) denotes the probability of a satellite i being infected with the epidemic pk at time t 1, once infected with epidemic pl at time t and migrating from an originating hub mo to a destination hub md.
6 SIMULATION AND ANALYTICAL RESULTS

We argue that a good context dissemination scheme reach nodes 1) with a low number of transmissions (efficiency), 2) with the most detailed context, and 3) with valid context w.r.t. time-space constraints (i.e., maximizing the remaining time until the obtained context turns obsolete). High values of the relaying probabilities, B _hh; _hs; _sh_>, can guarantee high probability of having the information reach a certain node but cause, as a side effect, flooding of the network and a high overhead which may jeopardize the
provision of other services. The proposed context model in Section 4 and the information dissemination model in Section 5 are compared against simulations and certain models in the literature.
6.1 Simulation Setting The adopted real-world networks for experimentation, such as social networks [18], are better modeled by disordered networks, e.g., small-world (SW) networks [19] and scale-free networks [34], than regular lattices [20]. Without loss of generality, CCA hubs can form a social network which is quite similar to the discussed scenario in our example (see Section 2). In order to evaluate the analytical SaIS model, we use a small-world network of M hubs with rewiring probability r and S satellites in total. The probability r denotes the probability that a hub gets rewired in a SWnetwork. In such network, a part of hubs, say out of M hubs, has the capability of inferring and disseminating context to the rest of the group. We call such hubs reasoners. The mean degree of connectivity of reasoners, w , plays significant role in the spreading of context, especially when context is temporally constrained. w denotes the probability of a reasoner to be connected with another reasoner. The coverage of the most detailed context pK across the network (i.e., the portion of nodes being infected with pK) indicates that the steady state of SaIS has been reached, i.e., all nodes converge to a common view of the surrounding environment. We consider the pieces of context pi; i 0; . . .;KK 3 that map to our example in Section 2 and form the hierarchy H fp3 _ p2; p2 _ p1; p1 _ p0g. Let nit 2 0; 1_ be the percentage of nodes that are infected with pi; i 0; . . . ; 3. The hubs and satellites are randomly chosen with initial probabilities of infection Pfxi;m0 pkg_T k0;...;K and Pfym0pkg_> k0;...;K; i1; . . . ; Sm0;m1; . . .;M, respectively, corresponding to a state pk. We set Pfx0 pkg < Pfx0 plg and Pfy0 pkg < Pfy0 plg with pk _ pl. This denotes that at the beginning of the process a small number of hubs and satellites are infected with strong epidemics. Furthermore, the mobility behavior of the satellites is represented by the probability of moving q as in (3). Satellites are gathered to the most popular hubs, thus forming an inhomogeneous network. A satellite moves to a randomly selected destination hub such that the probability of the pair of the originating and destination hub mo;md is selected with probability Pfho; hd mo;mdg as in (6). The simulation parameters are listed in Table 2. We use the MATLAB framework for all 10,000 simulations. In this section, we: . compare the performance of the proposed dissemination model against simulation results w.r.t. epidemic prevalence and decay for all epidemic levels, . examine how mobility of satellites influences the dissemination of pK across the network, . estimate the percentage nK with pK, and how it depends on (percentage of reasoners) and w (reasoners mean degree of connectivity), . compare SaIS with the models in [24] and [26], and the Gossip algorithm (in [7]) w.r.t. certain efficiency metrics in small-world and scale-free networks. Such models are discussed in Section 3. 6.2 Model Performance
At first we validate the theoretical results for SaIS through simulations. Fig. 3 depicts the performance of SaIS w.r.t. simulations in the static case for K 3. We use small-world networks with r 0:5, M 100 hubs, S 1;000 satellites (initially, satellites are randomly attached to their hubs), B 0:1; 0:1; 0:1_>, full cure rates _10 _20 _30 0:01, partial cure rates _21 _32 0:01, _31 0:001, and the initial probabilities of infection for both satellites and hubs are 0:9; 0:06; 0:03; 0:01_.We observe the evolution of n1t, n2t, and n3t. At the beginning of the simulation, it is observed that n3t < n2t < n1t. That is because, nodes first circulate p1, which can be further substituted by a refinement or be used for inferring more detailed context. In the long run, more knowledge is inferred across the network, thus, n2t dominates n1t. Finally, the most detailed context p3 is received by the majority of nodes, thus, signaling the steady state of the network. At the aggravation
TABLE 2 Simulation Parameters
phase (see Fig. 3), a node replaces pk by pm once pm _ pk. The difference between the theoretical and simulation curves is due to the value of the relaying probabilities. Moreover, the higher the value of infection rates is, the higher the network coverage is. Fig. 4 shows the propagation of p3 for B 0:05; 0:05; 0:05_>, B 0:1; 0:1; 0:1_>, B 0:2; 0:2; 0:2_> and B 0:5; 0:5; 0:5_> (from left to right). Now, we compare the performance of SaIS seen through simulations w.r.t. the decay of the epidemic process. We retain the same parameter setting as above (i.e., B 0:1; 0:1; 0:1_>) and replace the full/partial cure rates with _ij 0:1; i 1; . . . ; 3; j 0; . . . ; i. Fig. 5 shows the analytical and simulation curves in this case. It is observed that, the curves for SaIS and the simulations nearly coincide with each other, and, the stronger the epidemic is the faster it fades out with time.
6.2.1 The Impact of Mobility We examine how the mobility of satellites impacts the dissemination of the most detailed context pK. The performance of SaIS in light of mobility is studied once the distribution of the satellites in each hub remains constant as time progresses. At the beginning of the dissemination process, each hub has equal number of attached satellites (Smt 0 S M ;m 1; . . .;M). As satellites move to different islands, the most popular hubs gather a larger number of satellites than other hubs. After a certain number of relocations, each hub gathers a constant number of satellites. Fig. 6 depicts the predicted mean number of satellites Sm attached to hubs derived from SaIS and simulations having 0:02 and w 0:2. We observe an 45.84 percent increase in Sm for reasoners that infer pK and a 33.75 percent decrease in Sm for reasoners that infer less detailed context than pK. We calculate the probability Pfxi;mt pKg; i 1; . . . ; Smt;m 1; . . .;M denoting the infection of satellites by pK in the network. Initially, we set the state of all nodes (satellites and hubs) to p0 apart from a randomly selected hub infected with pK, i.e., M1. In addition, we retain the same network parameters setting (r; S;M; ;w ) and consider the probability of movement q 2 0; 1_ of the satellites in (3). A high value of q denotes that, in each step, it is likely that satellites move to different islands. The analytical and simulation values for Pfxi;mt pKg
are depicted in Fig. 7 for various values of q and 10,000 simulation runs. We can observe the impact of the mobility of the satellites: once a hub is infected with pK then the corresponding attached satellites have the opportunity to receive pK with probability bhs. In addition, satellites move between islands, thus, creating short-cuts, and relay pK sooner than in the static case. More interestingly, high probability values are observed even for low values of q and
Fig. 4. Simulation and analytical results for epidemic prevalence in disseminating p3 (static case) for various infection rates. Fig. 5. Simulation and analytical results for epidemic decay of p1, p2, and p3 epidemics. Fig. 6. The mean number of the attached satellites per hub versus time with 0:02 and w 0:2. Fig. 3. Simulation and analytical results for epidemic prevalence in disseminating p1, p2, and p3 (static case).
the rate of the increase on Pfxi;mt pKg decreases with the high mobility of satellites. Moreover, the mobility of satellites with relatively low infection rates _sh; _hs expedite context dissemination significantly. For a (low) value of q 0:05, we obtain an increase of 209.22 percent in Pfxi;mt pKg in the static case having the same _sh _hs 0:1. For q 1, we obtain an increase of 285.71 percent in Pfxi;mt pKg in the static case having the same _sh _hs 0:1 (similarly, 316.66 percent with _sh _hs 0:2). The coverage of a network with pK depends also on the percentage of the reasoners and their degree of connectivity w . Let q 0:5 and 0:05. We examine how the probability Pfxi;mt pKg is influenced with the value of w . A high value of w means that the reasoners are very popular and tightly interconnected than other hubs in the network. Hence, a hive number of satellites gradually visit them in order to receive detailed context. Fig. 8 depicts that satellites are much more likely to receive pK in a short time from the beginning of the dissemination process as long as they visit very popular hubs with high w . In addition, the probability Pfxi;mt pKg depends on the connectivity w . For slightly dense hubs (w 0:2) and q 0:5 we obtain a mean increase of 126 percent in Pfxi;mt pKg when ranges from 0:02 to 0:08.
6.3 Comparative Assessment 6.3.1 Performance Metrics We investigate the capability of the SaIS to efficiently disseminate pK. That is, we focus on the percentage of nodes that were infected with pK but also on the lowest cost in terms of network load. We define efficiency e 2 0; 1_ as the portion of nodes nk infected with some pk; 0 < k K out of the number of contacts (for message passing) c0k of those nodes to transit from p0 to all pk, that is, e 1 K XK
k1
nk c0k : 25 Ideally, the efficiency e should approach unity. Since the infected nodes may be reinfected by stronger epidemics, we define l-level efficiency el 2 0; 1_ as the portion of the infected nodes nlkt transiting from pl to all pk; 0 < k K; 0 < l < k out of the number of contacts clk, that is, el 1
Kl
PK
kl1 nlk clk
. The aggravation efficiency is defined as the sum of all el for 0<l K, that is, _e 1 K XK
l1
el: 26 Similarly, _e should also have a high value indicating that nodes efficiently transit between middle-level epidemics in order to be infected with pK. _e denotes the efficiency of a multiepidemic algorithm in disseminating information once the nodes are, at least, infected with some epidemic. Note that, if we set l 0 in el, we obtain that e0 e, and for K 1, we obtain 1-level efficiency; the known efficiency metric in information dissemination. Besides the e and _e metrics for comparative performance assessment, additional performance criteria that we adopt for assessing the various models are: time average of global infection status with epidemic of any level k > 0 and efficient information delivery in terms of information content transmission. The time average of global infection status metric _t 2 0; 1_ denotes the time average percentage of the number of nodes nkt; 1 k K that have been infected by k-level epidemic up to time t. Since, n0t1 PK k1 nkt we obtain, _T t 1 1
t
Pt
0n0,
where T is the total simulation time. For an information dissemination model, we assess the capability of infecting a large number of nodes with epidemic(s) in a given time horizon T. More specifically, the _T t metric for a dissemination model expresses the rate of infection up to time t; high (low) rate is obtained once _t is close to unity (zero). In order to define the efficient information delivery metric, we have to determine whenever a message is meaningful for a recipient node or not. Consider an infected node i by an epidemic of level l > 0, which at time t sends an l-level virus (message) to its neighboring node j 2 Vit. The message is considered meaningful for the node j if the node j is either susceptible or infected by epidemic of level 0 < u < l at time t. Otherwise, the message is meaningless for the node j. Let mulT be the total mean number of the meaningful messages over T that are sent for transmuting epidemic of level u to level l in time horizon T. Hence, efficient information exchange involves the dissemination of as many meaningful messages as possible. We define the
Fig. 7. Analysis and simulation results w.r.t. the mobility probability (q) with 0:02, w 0:2, B 0:1; 0:1; 0:1_>. Fig. 8. The probability Pfxi;mt pKg with w for 0:05; q 0:5.
efficient information delivery metric _K 2 0; 1_ for delivering epidemic of infection level K through intermediate multiple infections of lower epidemic levels as _K 1 K XK
l1
1 l Xl
u0
mulT culT ! ; where culT is the total number of contacts (transmission of meaningful and meaningless messages) that are used for transiting nodes from pu to pl; 0 < u K; 0 < l < u epidemic within time horizon T. The most important difference of the efficiency metrics (eK and e) with the _T t and _K metrics is that the latter metrics examine the capability of a model in propagating as meaningful information content as possible within a given time horizon. On the other hand, the efficiency metric generally counts the portion of the infected nodes out of the required contacts irrespective of the absolute coverage of infected nodes ( PK k1 nkt out of M S) and the number of meaningless information delivery messages (cklt mklt). Therefore, we have to consider all the metrics in order to objectively compare the dissemination models. We, then, define a holistic metric _ 2 0; 1_ for a dissemination model, which indicates time average coverage of infected nodes along with efficient information content transmission within time horizon T, thus, _ !_T T 1 !_K: 27 The ! 2 0; 1_ weight balances the importance between _T T and _K. A value of _ close to unity denotes that the dissemination model can infect the entire network with epidemic of the highest level in an efficient manner early in time.
6.3.2 Comparison with Other Models The considered context dissemination models for comparison are the model in [24] (model A), which adopts the SI epidemic model, the model in [26] (model B), which adopts the flooding scheme, and the underlying Gossip model in [7]. The technical details of the models are discussed in prior work Section 3. The model in [24] does not consider intermediate levels of context, thus, the nodes can only receive 1-level context. In this case we obtain _ij 0; 8i; j and _K _1, thus, the _e metric is not applicable for comparison. The model in [26] assumes dissemination of a context value p with a corresponding reliability indicator Ip. The dissemination model assumes B 1; 1; 1_, i.e., implements flooding. This means that a node communicates deterministically with all neighboring nodes. A node in [26] replaces a p value once Ip is below a threshold. In our abstraction Ip has the semantics: if Ip > Iq then p _ q, meaning that p is more reliable than q. Finally, we experiment with the Gossip algorithm, in which each node picks, according to some underlying deterministic or random rule, another neighboring node and exchanges information (a.k.a. rumor) with it. Two basic schemes are discussed in the literature for neighbor selection: the uniform gossip, in which each node chooses to communicate with a randomly chosen node at each step [35], and the standard gossip, in which a node picks, according to a probabilistic distribution, one of its immediate neighbors [36], [37], and [7]. We compare the SaIS model with the local
(uniform) Gossip model, in which a node i uniformly chooses to communicate with one neighboring node j2Vit. We compare all models w.r.t. the _ (! 0:5) and efficiency metrics (e; _e), while hub nodes form small-world and scale-free networks (also investigated in [21] for static networks). We adopt the Barabasi-Albert (BA) [34] algorithm for scale-free networks which is based on preferential attachment. The preferential attachment characteristic denotes that the more connected a node is, the more likely it is to receive new contacts. The Phwi distribution resulting from BA is of the form Phwi hwi_ with exponent _ usually between 2 and 3; hwi is the nodal degree. We construct the BA network of hubs starting with M0 10 (in the initialization phase of the BA algorithm) until M 100 hubs. For the simulations of the SaIS model we have _ij 0:001; i 1; . . . ; 3; j 0; . . . ; i. Table 3 depicts the efficiency e and aggravation efficiency _e metrics with different values of B (in case of SaIS and model A) for SW (r 0:5) and BA networks. SaIS assumes very high e value (e 0:86) for infection rates B 0:01; 0:01; 0:01_>, which clearly indicates that even with low infection rates, SaIS is an efficient way of covering the entire network. This is due to the exploitation of the semantic dependencies among pieces of context; nodes share intermediate context levels in order to achieve the higher level of context. In local Gossip and model A, we obtain similar e values to SaIS with infection rate 0.1 and 0.01, respectively. On the other hand, the model B demonstrates the lowest e value (e 0:12). When dealing with intermediate context levels, a high value of _e (_e 0:56) is obtained by SaIS, while in model B and local Gossip we obtain _e 0:003 and _e 0:01, respectively. This indicates that SaIS is more efficient in disseminating more detailed context to nodes, which have recently received context, than the other schemes. All models perform better in BA rather in SW networks w.r.t. efficiency, except for local Gossip. We obtain 6, 164, and 133 percent increase in the e value for SaIS (infection rate 0.01), model A (infection rates 0.01), model B, and 22 percent decrease in local Gossip models, respectively. The _e demonstrates quite similar behavior with e for all models in the BA network.
TABLE 3 Efficiency (e and _e) in SW and BA Networks
We now assess the performance for all models w.r.t. the holistic metric _, which is applicable to all models. _ indicates time average coverage of infected nodes along with efficient information content transmission with T 100. Fig. 9 plots the _ value for the models versus _, such as B _sh _; _hs _; _hh __>, in SW networks of hubs. In Fig. 9a, SaIS assumes a constant high _ value for _ > 0:1 (_ 0:8), while the B model assumes the higher _ value (0.67) compared to local Gossip, A model, and to SaIS for _ < 0:05. Moreover, local Gossip performs better than the A model for _ < 0:4. For _ _ 0:4, model A and local Gossip demonstrate similar behavior in context dissemination. Fig. 9b corresponds to a totally random SW network. In this case, local Gossip demonstrates the lowest _ value, while we obtain similar behavior in the _ metric for SaIS. Finally, we obtain a rather constant _ for the A model close to 0.55 w.r.t. _. Finally, Fig. 10 plots the _ value for all models versus infection rate for BA network of hubs. In this network, we
obtain the maximum _ value (0.78) for SaIS with _ > 0:3. The B model in BA network assumes similar behavior as in SW networks, while local Gossip assumes 20 percent decrease in _. On the other hand, one can observe 21.7 percent increase in _ values of the A model compared to local Gossip, for _ > 0:15. Overall, we can observe that the proposed SaIS model, which exploits semantic dependencies among pieces of context, behaves more efficiently than the considered models A, B, and local Gossip. Based on the _ and e metrics, we can conclude on the superiority of the SaIS model for context dissemination in terms of infection speed and efficiency for both SW and BA network types.
7 CONCLUSIONS
We introduce an information model for representing context by adopting hierarchical knowledge representation, and provide the corresponding reasoning process for inferring new context. In addition, we propose an analytical information dissemination model based on the epidemic spreading (SaIS). The context model exploits information specialization relations for inferring new context, thus, expanding the knowledge across a mobile network. We also use the analogy between context and epidemic and extend the monoepidemic SIS model to the multiepidemic SaIS model, in the sense that a node can be reinfected by a stronger epidemic. A strong epidemic denotes that each node can be reinfected with more detailed context than that it has recently received. In the monoepidemic spreading (i.e., with no information specialization), we obtain the analytical model studied in [3]. In the generalized SaIS model, each epidemic has different spreading behavior which affects the entire information dissemination process. Moreover, we study the impact of the node mobility in the spreading process speed. The context model and the SaIS model are assessed through analysis and extensive simulations in satellite-hub networks. We also provide a comparative assessment with other models in the literature for SW and BA networks finding that SaIS is efficient in the multiepidemic-based context dissemination. Future work concerns both the context and SaIS models. We focus on the stability of the context model in recognizing and handling dissemination of contradictory pieces of information. Moreover, the impact of a fully mobility pattern of nodes as long as the infection between distant groups is another interesting area we are currently working on.
REFERENCES
[1] B. Schilit and M. Theimer, Disseminating Active Map Information to Mobile Hosts, IEEE Network, vol. 8, no. 5, pp. 22-32, Sept. 1994. [2] A.-M. Kermarrec, L. Massoulie, and A. Ganesh, Probabilistic Reliable Dissemination in Large-Scale Systems, IEEE Trans. Parallel and Distributed Systems, vol. 14, no. 3, pp. 248-258, Mar. 2003. [3] C. Zesheng and J. Chuanyi, Spatial-Temporal Modeling of Malware Propagation in Networks, IEEE Trans. Neural Networks, vol. 16, no. 5, pp. 1291-1303, Sept. 2005. [4] J. Kephart and S. White, Directed-Graph Epidemiological Models of Computer Viruses, Proc. IEEE Research in Security and Privacy pp. 343-359, 1991. [5] A. Dey, Understanding and Using Context, Personal and Ubiquitous Computing J., vol. 5, no. 1, pp. 4-7, 2001. [6] A. Datta, S. Quarteroni, and K. Aberer, Autonomous Gossiping: A Self-Organizing Epidemic Algorithm for Selective Information Dissemination in Wireless Mobile Ad-Hoc Networks, Proc. Intl Conf. Semantics of a Networked World, pp. 126-143, 2004. [7] D. Kempe, J. Kleinberg, and A. Demers, Spatial Gossip and
Resource Location Protocols, J. ACM, vol. 51, no. 6, pp. 943-967, Nov. 2004. [8] A. Ku pper, Location-Based Services: Fundamentals and Operation. John Wiley & Sons, Oct. 2005. [9] V. Papataxiarhis, V. Riga, V. Nomikos, O. Sekkas, K. Kolomvatsos, V. Tsetsos, P. Papageorgas, S. Vourakis, S. Hadjiefthymiades, and G. Kouroupetroglou, MNISIKLIS: Indoor Location Based Services for All, Proc. Fifth Intl Symp. Location Based Services and TeleCartography, Nov. 2008. [10] P. Brezillon, Individual and Team Contexts in a Design Process, Proc. IEEE Intl Conf. System Sciences, 2003.
Fig. 9. The _ value for SW networks with (a) r 0:5, and (b) r 1. Fig. 10. The _ value for the dissemination models in BA network. [11] M. Borges, P. Brezillon, J. Pino, and J. Pomerol, Groupware System Design and the Context Concept, Proc. Computer Supported Cooperative Work in Design (CSCWD 05), pp. 45-54, 2005. [12] C. Anagnostopoulos, A. Tsounis, and S. Hadjiefthymiades, Context Awareness in Mobile Computing Environments, Wireless Personal Comm. J., vol. 2, no. 3, pp. 454-464, Aug. 2007. [13] P. Brezillon and J. Pomerol, Contextual Knowledge Sharing and Cooperation in Intelligent Assistant Systems, Le Travail Humain, vol. 62, no. 3, pp. 223-246, 1999. [14] C. Anagnostopoulos, P. Passias, and S. Hadjiefthymiades, A Framework for Imprecise Context Reasoning, Proc. IEEE Intl Conf. Pervasive Services, pp. 181-184, 2007. [15] S. Hamed, G. Bianconi, and M. Marsili, Scale-Free Networks with an Exponent Less than Two, Physical Rev., vol. E 73, p. 046113, 2006. [16] C. Anagnostopoulos, E. Zervas, and S. Hadjiefthymiades, An Epidemiological Model for Semantics Dissemination, Proc. Third ACM Intl Conf. Mobile Multimedia Comm. (MobiMedia 07), vol. 329, 2007. [17] J. Mickens and B. Noble, Modeling Epidemic Spreading in Mobile Environments, Proc. ACM Wireless Security, pp. 77-86, 2005. [18] S. Wasserman and K. Faust, Social Network Analysis. Cambridge Univ., 1994. [19] D. Watts, Small Worlds: The Dynamics of Networks between Order and Randomness. Princeton Univ., 1999. [20] M. Newman, The Structure of Scientific Collaboration Networks, Proc. Natl Academy of Sciences of USA, vol. 98, no. 2, pp. 404-409, 2001. [21] M. Boguna, R. Pastor-Satorras, and A. Vespignani, Epidemic Spreading in Complex Networks with Degree Correlations, Proc. SITGES Conf. Statistical Mechanics of Complex Networks, pp. 127-147, 2003. [22] Y. Wang, D. Chakrabarti, C. Wang, and C. Faloutsos, Epidemic Spreading in Real Networks: An Eigenvalue Viewpoint , Proc. 22nd Intl Symp. Reliable Distributed Computing, pp. 25-34, 2003. [23] M. Akdere, C. Cagatay, O. Gerdaneri, I. Korpeoglu, O. Ulusoy, and U. Cetintemel, A Comparison of Epidemic Algorithms in Wireless Sensor Networks, Elsevier Computer Comm. J., vol. 29, no. 13, pp. 2450-2557, Apr. 2006. [24] A. Khelil, C. Becker, J. Tian, and K. Rothermel, An Epidemic Model for Information Diffusion in MANETs, Proc. ACM Workshop Modeling Analysis and Simulation of Wireless and Mobile Systems, pp. 54-60, 2002. [25] M. Faloutsos, P. Faloutsos, and C. Faloutsos, On Power-Law Relationships of the Internet Topology, Proc. SIGCOMM, pp. 251262, 1999. [26] J. Mantyjarvi, P. Huuskonen, and J. Himberg, Collaborative Context Determination to Support Mobile Terminal Applications, IEEE Wireless Comm. Magazine, vol. 9, no. 5, pp. 39-45, Oct. 2002. [27] A. Ganesh, L. Massoulie, and D. Towsley, The Effect of Network Topology on the Spread of Epidemics, Proc. IEEE INFOCOM, pp. 1455-1466, 2005. [28] C. Anagnostopoulos and S. Hadjiefthymiades, Enhancing SituationAware Systems through Imprecise Reasoning, IEEE Trans. Mobile Computing, vol. 7, no. 10, pp. 1153-1168, Oct. 2008. [29] V. Tsetsos and S. Hadjiefthymiades, An Innovative Architecture for Context Foraging, Proc. Intl ACM Workshop Data Eng. for Wireless and Mobile Access (MobiDE in conjunction with SIGMOD/ PODS 09), 2009.
[30] M. Musolesi and C. Mascolo, Controlled Epidemic-Style Dissemination Middleware for Mobile Ad Hoc Networks, Proc. Third Ann. Intl Conf. Mobile and Ubiquitous Systems, 2006. [31] G. Chen and D. Kotz, Policy-Driven Data Dissemination for Context-Aware Applications, Proc. Third IEEE Intl Conf. Pervasive Computing and Comm., pp. 283-289, 2005. [32] D. Ejigu, M. Scuturici, and L. Brunie, CoCA: A Collaborative Context-Aware Service Platform for Pervasive Computing, Proc. Fourth Intl Conf. Information Technology (ITNG 07), pp. 297-302, 2007. [33] A. Salkham, R. Cunningham, A. Senart, and V. Cahill, A Taxonomy of Collaborative Context-Aware Systems, Proc. CAISE Workshop Ubiquitous Mobile Information and Collaboration Systems (UMICS 06), pp. 899-911, 2006. [34] R. Albert and A-L. Barabasi, Statistical Mechanics of Complex Networks, Rev. of Modern Physics, vol. 74, no. 1, pp. 47-94, Jan. 2002, doi:10.1103/RevModPhys.74.47. [35] U. Feige, D. Peleg, P. Raghavan, and E. Upfal, Randomized Broadcast in Networks, Proc. Intl Symp. Algorithms, pp. 128-137, 1990. [36] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, Randomized Gossip Algorithms, IEEE Trans. Information Theory, vol. 52, no. 6, pp. 2508-2530, June 2006. [37] L. Xiao, S. Boyd, and S. Lall, A Space-Time Diffusion Scheme for Peer-to-Peer Least-Squares Estimation, Proc. Fifth Intl Symp. Information Processing in Sensor Networks, pp. 168-176, 2006. Christos Anagnostopoulos received the BSc degree in computer science from the Department of Informatics and Telecommunications at the National & Kapodistrian University of Athens (NKUA), Greece, in 2001 and the MSc degree in computer science (advanced information systems) from the same department in 2003 and the PhD degree in modeling mobile and distributed computing systems in 2008 from NKUA. His research interest is focused on mobile and distributed computing systems, context-aware computing, semantic web, and ontological engineering. He had also participated in projects realized in the context of EU programs. He is a member of the IEEE Computer Society. Stathes Hadjiefthymiades received the BSc (with honors) degree in computer science from the Department of Informatics at the University of Athens, Greece, in 1993 and the MSc (with honors) degree in computer science (advanced information systems) from the same department in 1996. In 1999, he received the PhD degree from the Department of Informatics and Telecommunications, University of Athens. In June 2002, he received a joint engineering and economics MSc degree from the National Technical University of Athens. Since 1992, he was with the Greek consulting firm Advanced Services Group, Ltd. (ASG, Ltd.), where he was extensively involved in the analysis and specification of information systems as well as the design and implementation of telematic applications. Since 1995, he has been a member of the Communication Networks Laboratory at the University of Athens. He has participated in numerous projects realized in the context of EU programs (ACTS, ORA, TAP, INFO2000, IST) as well as national initiatives (Telematique, RETEX, Mentor). From September 2001-July 2002, he served as a visiting assistant professor in the Department of Information and Communication Systems Engineering at the University of Aegean. He joined the faculty of the Hellenic Open University, Patras, Greece, in summer 2002 as an assistant professor of telecommunications and computer networks. Since December 2003, he has belonged to the faculty of the Department of Informatics and Telecommunications, University of Athens, where he is presently an assistant professor of large-scale software systems. His research interests are in the area of wireless/mobile/pervasive computing and networked multimedia applications. He is the author of more than 100 publications in international scientific journals, conferences, and books in the above areas. He is a member of the IEEE Computer Society. Evangelos Zervas received the BSc degree in electrical engineering from the National Technical University of Athens, Greece, in 1986, the MSc degree from Northeastern University
(NEU), Boston, Massachusetts, in 1989, and the PhD degree in communications and digital processing from NEU in 1993. His is now a professor in the Department of Electronics, Technological Educational Institutions, Athens, Greece. His research interests are in the area of wireless sensor networks, mobile computing, and sensor data fusion. He is a member of the IEEE Computer Society. . For more information on this or any other computing topic, please visit our Digital Library at www.computer.org/publications/dlib.

Information Dissemination Between Mobile

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Information Dissemination Between Mobile

Uploaded by

Copyright:

Available Formats

Information Dissemination between Mobile Nodes for Collaborative Context Awareness

on Context-Awareness (CA), which is the

1. The spreading of an independent piece of information.

4 CONTEXTUAL INFORMATION MODEL

5 INFORMATION DISSEMINATION MODEL

TABLE 1 Algorithm: Node Reasoner

Rl;k _ PfYVm;h t yVm;h tg YSmt

Rl;k _ PfYVmo;h t yVmo;h tg _ SmYo t1

Rl;k _ PfYVmd;h t yVmd;h tg _ Y Smd t

6 SIMULATION AND ANALYTICAL RESULTS

TABLE 2 Simulation Parameters

TABLE 3 Efficiency (e and _e) in SW and BA Networks

You might also like