Esprit "IT for Mobility" 26900
|
Esprit "IT for Mobility" 26900 |
|
|
|
|
|
|
| Pre-processing | |
| Before we start clustering
the sensor data together, we improve the input of the system by performing
small, fast and often very simple precalculations we call cues.
The frequency of the light, or cycles of motion of the accelerometers are
examples of cues. This way, the learning algorithm has better (more obvious)
values and can converge faster to a near-100% recognition.
A cue is a logical sensor that represents information on a more abstract level. A cue can be represented as symbolic or sub-symbolic value. There are actually two general methods to create a cue. First, the cue can be calculated by applying a pre-processing step to the data from the physical sensor. In this case a cue is always based on a single sensor, whereas the data provided by a physical sensor can be used to generate a number of different cues. Second, already symbolic values, such as time or scheduled date, are considered as cue. An actual context – regarded as a logical sensor - could become a cue in the next recognition cycle as well. |
The information that we get from the sensors is far more powerful than one might expect on first glance. A light sensor, for instance, does not really compare to human vision. Most people underestimate it by thinking it is just usable for measuring the intensity of light, but it is able to send its values very rapidly (hundreds of times per second). Electric lights (normal light bulb, TL light, halogen light, …) have a certain frequency which is too fast to see for our eyes. The light sensor is thus able to see artificial lights go on and off, which is something that humans (and most animals) can not do. This makes a light sensor very useful in distinguishing various sources of light. |
| The learning architecture: Clustering, Classification & Supervision | |
 A
context (description) is calculated from the information delivered by cues.
In general, a context is based on a number of cues, but it is also possible
that a context can be derived from a single cue. A context is a symbolic
value that describes the current situational context.
The learning architecture is described in "Online adaptive context awareness". Basically, the idea is to cluster the input signals (cues) with a neural network called the 'Kohonen Self-Organizing Map' (KSOM). The outcome of this clustering is then labeled or classified, after which it is checked by a probabilistic finite state machine. |
|
| last update: 27/06/2000 by Kristof Van Laerhoven | the previous sites can still be found here and here |