We consider situation-aware mobile services as applications that could bene t considerably from advances in stream reasoning technology. Our initial realization of a situation-aware mobile service infrastructure combines streams of context information gained from mobile sensors and various Web resources and applies expressive reasoning. With the availability of stream reasoners, this distributed computing infrastructure could be made considerably simpler and more powerful. Recent ndings in applying and evaluating existing reasoning engines are taken into account to highlight essential requirements and potential directions for future research on stream reasoning.
In the recent past, the mobile phone has seen a dramatic shift in its perception from a plain communication device to an intelligent source of relevant information for the very moment. Location-based services, rst introduced a couple of years ago, paved the way for much more useful mobile services that assist the user in his spatial surrounding. Foreseen as the successor of these mere location centric services, context-aware applications are meant to enable a multiplicity of intelligent agents, capable to interpret heterogenous data sources and thus enabling truly intelligent services.
With IYOUIT1, a framework for novel mobile applications, we recently
introduced a mobile community service in the eld of context-awareness [
IYOUIT has been realized as a light-weight mobile client application that
connects to a distributed server infrastructure [
We see a high potential in the application of semantic technologies, in
particular the deployment of the Web Ontology Language (OWL) [
To deploy a situation-aware service infrastructure with already existing semantic technology involves a couple of architectural design decisions, as depicted on the left-hand side of Figure 1. With mobile handsets becoming more and more powerful in terms of the number of built in sensors, the client application has been designed to gather and combine various sensor outputs. For further processing, lower-level sensor data is sent (approximately every 2 to 20 minutes) to the network infrastructure via a permanent mobile data connection. The realtime constraints for the situation computation limit the maximum classi cation time to not more than a second.
Due to the limitations of current reasoning engines in e ciently dealing with
large data sets in expressive knowledge representations, further abstractions have
to be derived before the actual reasoning can be applied. In our current
architecture, the client server communication follows a standard request reply pattern,
in which the mobile client triggers the actual situation detection as soon as any
of the sensory inputs registers a change in its state. However, only two minimal
snapshots can be compared to decide whether or not a new computation of the
current situation is meaningful or not. With regard to the fact that the
underlying data is often noisy, uncertain or even inconsistent, this decision is hard to
make in the given setting. For example, in case no GPS x is available, IYOUIT
computes the phone's position by mapping the connected base station identi er
to an estimation of its location (latitude, longitude and range) based on previous
measurements that have been stored in a DB [
Here, stream reasoning could improve the detection of meaningful classi cation attempts in relocating the event driven evaluation of state changes to the networked components (cf. right-hand side of Figure 1). With far more com
Reasoner
Sensors Mobile handset
Mobile network Network infrastructure
SR SR SR SR
CP
SR: StreamReasoner CP: ContextProvider SP: SituationProvider External data source publish & subscribe + n to n i taup itago om reg tac ag l e D CP
CP
External data source
SR
SP
DB puting power available and the possibility to e ciently persist data, larger timeframes could be taken into consideration. For example, the error rate of detecting a driving event based on the current speed computed by taking the delta of two location estimations is far too high, in case no GPS speed is available. If a series of geo coordinates and their estimated ranges are considered, a driving event can be detected reliably with simple geometrical calculations that assume a directed move. Furthermore, a subscription based model of interconnected components might distribute the previously centralized reasoning task. 3
To nd the optimal selection of an OWL reasoner and service interface for our
application that comply with the real-time constraints we conducted extensive
performance evaluations with knowledge bases of di erent size and complexity
[
It became also apparent in our study [
Most OWL reasoning algorithms have been designed and optimized for relative static knowledge bases.
For applications with frequently changing data like expressive syndication
[
Todays incremental query answering algorithms are limited to theories with only simple roles whereas our examples rely on transitive rules. Here, the incremental reasoning algorithms under syntactic addition and removal of extensional knowledge implemented as part of the incremental consistency checker of newer Pellet version should be able to considerable speed up our use case. However, we could not detect any speci c performance gain of this implementation. Still, in case the actual reasoning task can be computed very fast and the necessary axioms representing the changing data can be transmitted e ciently, the retrieval of large inference results back to the application accounts for a time critical task in the overall process. 4
We introduced an application scenario that could bene t from advances in the
eld of stream reasoning. Especially, the use of stream reasoners following the
revolutionary approach [
Our recent performance results clearly demonstrate the need for reasoning support in modular theories, fast standard axiom level communication protocols supporting retraction (such as OWLlink), and incremental inference algorithms as enabling technologies for stream reasoning. At the same time we want to
raise our concerns regarding the correctness of even more sophisticated inference engines, especially, if inference results are used in critical ways (e.g., for contextdependent access control mechanisms).