Context-aware systems must be able to deal with uncertain context information. We propose a generic context architecture and representation that incorporates the uncertainty of context elements in terms of upper and lower bounds of probabilities. It is shown how opinion nets can be used to reason with these upper and lower bound probabilities. In this way it is possible to combine ambiguous or con°icting context information that comes from di®erent sources. Moreover, information coming from di®erent sources can be combined with experience learned from the past in a clean way.
Pervasive systems that can adapt to changing environments and availability of resources must be aware of their context. These systems sense and react to context. Most systems make the assumption that the context they use is completely accurate. However, the information about context may not come from a reliable source, may be out dated, not available or may be erroneous. Firstly, a contextaware system senses its context via a network of sensors working together. The resolutions, accuracies and formats of these sensors can di®er from each other. The resulting sensed values can have con°icts and ambiguities. The second cause of uncertainty are the current limitations of the underlying reasoning systems that deduce high-level context information from low-level sensor data. Lastly, due to the asynchronicity of context acquisition and use of context we must deal with the imperfection and aging of the context information. A challenge for the development of real-life and commercial context-aware systems is therefore the ability to handle uncertain and ambiguous context information.
We propose a generic context architecture consisting of context synthesizers, providers and consumers. Context elements are represented as predicates, with which are associated upper and lower bound probabilities. Then opinion nets are used to reason with these probabilities. If the context comes from di®erent sources contradictions and ambiguities can arise. It is shown how opinion nets can resolve con°icts and ambiguities by combining several probabilistic inputs to a single output.
This paper is organized as follows. In section 2 the context architecture and representation are described. In section 3 we introduce opinion nets and show how they can be of great value to deal with uncertainty in context-aware systems. Section 4 gives an overview of related work and compares it with the presented approach. Finally, section 5 gives an outline of future work and draws conclusions. 2
The context architecture is a generic infrastructure inspired by Gaia [
The general uncertainty model is based on predicates representing context elements or facts with associated con¯dence values. The predicate name describes the context element. The arguments are mostly of the form subjectobject or subject-verb-object, e.g. location(John, in, room3, lower, upper) or activity(room7, conference, lower, upper). The con¯dence values of the predicates are expressed as upper and lower bounds of probabilities. Alternatively we could express con¯dence values as a probabilistic value together with an accuracy measurement of the probability. We will call the con¯dence value an opinion. 3
In a simple approach we could work with a singular probabilistic value to indicate
the frequency that a predicate is true. However, in opinion nets [
Opinions coming from di®erent sources can be tied together in several ways. The di®erent sources could for example be one or more context providers and context synthesizers. The combining of the inputs to one output can be done in a context provider, synthesizer or consumer, depending on the requirements and structure of the application. The inputs of a context provider or synthesizer can also be put together with historical information that is learned from the past. In that way history can be taken into account and easily incorporated in the opinion net approach.
The advantage of working with upper and lower bounds is that we can work with imprecise probabilities. When there is not enough information to give an exact probability, but if the system knows enough to say that the probability is de¯nitely between 0.4 and 0.7, we can capture this with opinions. Also, ambiguous information can be presented in a straightforward way. Lastly, con°icting information that comes from di®erent sources can be combined in a clean way by using opinion nets.
Figure 1 shows the di®erent opinions and how they are put together, for clarity only for the lower bounds. The opinions can be propagated through opinion nets. The boxes are called constraint boxes, and they can be and or or boxes. The following constraint equations govern the action of the or boxes. A and B represent inputs and A or B represents an output. Then, l(A), l(B) and l(A or B) are the lower bounds of the probabilities. Similarly u(A), u(B) and u(A or B) represent the upper bounds of the probabilities.
u(A) · u(AorB) l(A) ¸ l(AorB) ¡ u(B)
u(B) · u(AorB) l(B) ¸ l(AorB) ¡ u(A) u(AorB) · u(A) + u(B) l(AorB) ¸ max[l(A); l(B)] (1)
The equations for the and operator are similar. The combination of these assertions and boxes is an opinion net. An opinion net is thus a numeric constraint net in which it is possible to keep track of a conclusion's probability. 4
There exist several approaches for dealing with uncertain, ambiguous and
inconsistent context information. The Integrated Context Model proposed by Truong
[
Gaia [
Several techniques can be used to deal with imperfect context. However, Dey
and Manko® [
To validate our approach a simulation environment will be developed. Experiments with context information coming from di®erent sources have to be carried out. A test scenario can be the introduction of a vague concept like proximity as a context element. Context-aware systems such as a portable touristic city guide are location-aware and can suggest a tourist to visit a touristic attraction that is nearby. Proximity however is a subjective measure of distance depending on the context of the tourist. How close an attraction is depends amongst others on whether the tourist is on foot or by car, what his mood is, whether he is really interested to see the attraction and so on. Opinions concerning the proximity of a location that come from di®erent sources can be combined by an opinion net to a single output of a probability range.
Based on a context architecture and representation, we proposed the use of opinion nets to deal with uncertain, ambiguous and con°icting context information. This approach allows to resolve ambiguities and con°icts arising from information that comes from di®erent sources in a natural way. Moreover, the reasoning mechanism with opinion nets is general in the sense that it can be plugged in into every context system that represents information with probabilities and accuracies. We believe that further research will show the usefullness of opinion nets for reasoning with uncertain context information.