Title: Discovering relevant sensor data by Qanalysis
1Discovering relevant sensor data by Q-analysis
- Pejman Iravani
- Department of Design and Innovation
- The Open University
2Introduction
- The problem of adding new sensors
- In principle, adding sensors to a robot seems
beneficial as more information is given to the
robot. - But, this is not the case if the sensors added
provide irrelevant or redundant data. In which
case adding sensors has negative impact due to
the curse of dimensionality.
3Introduction
- Curse of dimensionality is related to the
exponential growth of a hyperspace as a function
of its dimensions.
Adding irrelevant sensors increases exponentially
the size of the state space and thus the
computational requirements of the system dealing
with such space.
4The feature selection problem
- Given a set of m features (robot sensors) select
a subset of n relevant features, where mgtn, such
that the subset provides the same or similar
information than the original set.
5The feature selection problem
- The relevance of the features is assessed in
relation to classification. - A feature is relevant if it provides useful
information for discriminating entities which
belong to different classes.
6Classification by Q-analysis
- Given a robot with some binary sensors.
- At any point in time a number of sensors will be
active (e.g. pressed bumpers), which are
represented by simplices s.
7Classification by Q-analysis
sltx1,x2,x3gt
sltx1,x2,x3,x4gt
- The dimension of the shared face is known as the
q-nearness of two simplices. - A fundamental hypothesis is that q-nearness is a
measure of structural similarity.
8Classification by Q-analysis
- The idea of q-nearness can be extended to a set
of simplices, in which case, their largest shared
face is known as hub.
Classification using Q-analysis is based on
finding relevant hubs with respect to each
class. The vertices that are not included in the
hubs can be considered as irrelevant.
9Feature selection by Q-analysis