Information Technology Reference
In-Depth Information
cases. This can lead to the situation that improving the situation at one point in
the road network will worsen the situation somewhere else. Simulation provides
a useful instrument in order to investigate different strategies and to find out
about effects in different settings. Nevertheless, it should be at least mentioned
that simulation does not come for free as a certain modeling effort has to be
considered and that simulation results are not always trusted.
We present an approach to learning dynamic adaptation strategies for simu-
lated trac scenarios. The basic idea is to extract features of the current trac
situation and to investigate the effects of various activities. The desired outcome
of this process is to identify a strategy which can dynamically adapt the behavior
of certain infrastructure elements (e.g., road segments) or to provide suggestions
to road users how they should behave in order to improve the situation. We ap-
ply supervised symbolic machine learning algorithms in order to find out about
useful rules. Having in mind situation descriptions and potential actions with
the aim to learn a successful strategy, reinforcement learning techniques seem to
be very well suited. However, in this work we have the additional requirement
that comprehensible rules should be generated which can be understood and
maybe even manually adapted by experts. We apply the approach to two sce-
narios. In the first scenario, it is learned what speed limit should be set on a road
depending on the actual situation. In the second scenario, a variable-message
sign (VMS) can force road users to change the lane in order to let cars enter the
motorway. A classifier is learned in which situations the VMS should be active.
Although it would also be possible to learn dynamic adaptation strategies for
road users by using their individual or even joint experiences for learning, we
focus on a central infrastructure-based setting in this paper.
2
Related Work
The combination of data mining and agent-based systems is discussed in recent
work. Baqueiro et al. present a study where they discuss two different approaches
for a combination of the two fields [1]: 1) Using data mining techniques for
investigation of agent-based modeling and simulation (ABMS) and 2) Utilizing
ABMS results in data mining. Many works have also been done in the field of
Adaptive Trac Control Systems (ATCS). For instance, Bull et al. [4] discuss
learning classifiers in order to improve trac-responsive signal control systems.
A review of ATCS is provided, e.g., in [16].
Gehrke and Wojtusiak [7] present an approach to react online on influences
from the environment (for example weather) in the context of route planning.
The approach tries to identify the best routes by taking into account the wetness
and the speed limits of the roads. Every truck is represented by an agent and
can dynamically react to new events. The (propositional) rule induction system
AQ21 has been used to learn prediction rules.
Bazzan et al. have investigated the adaptation of driver behavior and trac
light agents [3]. The motivation for their approach is an integrated consideration
of trac light control and route choices for drivers. The adaptation of both types
