بخش بندی شبکه های بیزی برای انتشار اطلاعات و آگاهی در محیط های آگاه از متن و همکاری با Bayeslets
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|29303||2014||18 صفحه PDF||سفارش دهید||10381 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Pervasive and Mobile Computing, Volume 12, June 2014, Pages 214–231
With ever smaller processors and ubiquitous Internet connectivity, the pervasive computing environments from Mark Weiser’s vision are coming closer. For their context-awareness, they will have to incorporate data from the abundance of sensors integrated in everyday life and to benefit from continuous machine-to-machine communications. Along with huge opportunities, this also poses problems: sensor measurements may conflict, processing times of logical and statistical reasoning algorithms increase non-deterministically polynomially or even exponentially, and wireless networks might become congested by the transmissions of all measurements. Bayesian networks are a good starting point for inference algorithms in pervasive computing, but still suffer from information overload in terms of network load and computation time. Thus, this work proposes to distribute processing with a modular Bayesian approach, thereby segmenting complex Bayesian networks. The introduced “Bayeslets” can be used to transmit and process only information which is valuable for its receiver. Two methods to measure the worth of information for the purpose of segmentation are presented and evaluated. As an example for a context-aware service, they are applied to a scenario from cooperative vehicular services, namely adaptive cruise control.
The added value of ubiquitous computing systems stems from the fact that participating devices have access not only to their own information, but also to the information of other relevant service providers, sensors or information consumers. The information received from heterogeneous and redundant sources is usually evaluated and fused with one’s own information to form new, more accurate, or more reliable knowledge . This inference on the available information is necessary for complex situation estimations e.g. to avoid collisions in road transportation. Information is exchanged via centralised architectures like the Internet, but also in ad-hoc networks like vehicle-to-vehicle (V2V) communication networks in safety-critical driver assistance systems  (cf. Fig. 1) or communicating personal smart spaces . In particular the ad-hoc networks in the latter situations are always wireless, may exist only for some moments and their capacity is limited. Depending on the current situation, sensor measurements can have a high importance to a node and its neighbours (e.g. to prevent a collision). However, the transmission of information over the wireless channel consumes bandwidth and thus prevents others from transmitting their–maybe more important–information. Whether information shall be exchanged hence becomes an information-theoretic problem which requires decision-making based on the information which is to be disseminated and the costs implied by the dissemination. The algorithms of Section 4 introduce concepts for adaptive information dissemination using a utility-based decision-making approach. Full-size image (50 K) Fig. 1. Vehicle-to-vehicle communication helps to avoid collisions in unclear situations . Figure options The information that would be available to be exchanged in such mobile and highly dynamic environments exceeds the network capacity. Further to network bandwidth being a critical resource, also the processing capacity of many mobile computers would not be able to cope with an overload of perhaps even redundant information. Imagine the following standard example of ubiquitous computing, from the domain of Unified Messaging: Calls shall only be forwarded to the recipient if he is available, i.e. in a position to answer the phone. This depends on his agenda, his location, his current activity, time, the persons in his surroundings, ambient context (e.g. noise level) and also the caller and his status (e.g. calls in emergency situations should be forwarded). In addition, most of the mentioned input context information depends again on many other context sources and sensors. All this information has to be collected and evaluated on the resource constrained mobile phone to infer the availability. The powerful methods for Bayesian networks are a good starting point for realising such context inference. They will be the basis for the approach presented in this paper. The NP-hard complexity  of probabilistic inference however entails the need for dedicated algorithms which are able to cope with high amounts of information. The problem of information overload hence has to be solved, to reduce both, network load and computation time. Our general idea for this solution is the introduction of mechanisms where information is assessed before transmission and processing. The presentation of such mechanisms is the objective of this paper, i.e. rather the contribution to solving the logical problem in computability theory than the tuning of performance. We propose the Bayeslet concept to segment Bayesian networks, allowing for distributed inference of the situation and reduced transmission needs, as well as the “rating” of information which can automate the decision of which information needs to be transmitted for an inference target. We demonstrate the applicability in cooperative driving assistance using V2V networks as a special case of context inference where it reduces the network traffic and computational burden for processors. After a review of the background theory and the related work in the following section, Section 3 introduces the Bayeslet concept, a way to bundle information for local value determination and evaluation. Section 4 describes the approaches to define the worth of information which are then applied to an example from collision avoidance in vehicle networks in Section 5. Finally, we close this paper with a conclusion and an outlook on how we will pursue and integrate this work in the future.
نتیجه گیری انگلیسی
This work has proposed the Bayeslet concept for structuring context inference networks which allows one to cope with very large amounts of information both from a computational and a remote communications point of view. Bayeslets represent enclosed and encapsulated knowledge domains that are, due to their size, easily manageable and processable and allow for distributed evaluation and selection of relevant information. The restrictions, reduced modelling flexibility and inference precision, have to be accepted in order to take advantage of tractable and bandwidth efficient context inference. The approach presented for the composition of Bayeslets is novel and formal. It concretises the rather vague concept of relevancy used in many definitions of context. For a given objective, the most relevant information is selected and thereby, depending on an application specific threshold, information transmission is limited. This reduces inference time and increases bandwidth efficiency. In the presented simplistic example, we were able to show that in many cases the connection of a second Bayeslet is not useful. We could reduce transmissions over the network by about 90% and reduce the inference time necessary to incorporate a Bayeslet to 50%. The determinator of the worth of information, utility, can be defined in various ways. Both approaches presented yield different results for the application example. This is due to the different objectives of both approaches. The objective of mutual information based on the entropy is always the reduction of uncertainty. The objective of the decision based approach, however, can be defined freely by the configuration of the utility node(s). Consequently, both methods are not mutually exclusive, but are valid alternatives which can be chosen depending on the current situation. The creator/owner of a Bayeslet has the option to define decision and utility nodes for a specific objective. If no objective is defined, the standard would be the uncertainty reduction. This can be realised with the mutual information based composition decision. In the future these results already confirmed in theory, have to be tested in real inter-vehicle communication. To this end, the Bayeslet management system has to be incorporated in the vehicle communication system. A further step should be the consideration of maximum response times. They could be realised e.g. with an adaptive cost function for the composition of Bayeslets. In general, the work on cost functions for composition still offers room for research. A range of requirements (network response time, speed of computation, adaptation to situation, comparability among different Bayeslets, etc.) can be defined which has to be fulfilled for all different domains causing composition costs, such as time, memory and communication costs.