Information Technology Reference
In-Depth Information
of behaviors is carried out locally, i.e., without central control. In the cases where
no fixed or greedy settings are used, the trac light agent adaptation is based on
Q-Learning and for the routing decision a probabilistic selection is performed in
dependence of previous experience. The authors report an improvement in travel
time and occupancy for the combined adaptation in comparison to situations
where only drivers or trac lights are adapted. In another article, Bazzan [2]
discusses opportunities for multiagent systems and learning in the context of
trac control with a focus on reinforcement learning approaches.
The approach of Fiosins et al. [6] addresses the detection of change points
for intelligent agents in city trac. They apply a bootstrapped cumulative sum
charts (CUSUM) test and a pairwise resampling test in order to detect change
points. In a case study, the approach is applied to a vehicle routing problem.
The authors report that the change point detection and re-routing decision could
reduce the travel times of the vehicles.
In our work, we decided to use a supervised learning setting especially tak-
ing into account decision rule or decision tree induction in order to meet the
requirement of getting comprehensible and manually adaptable results. In our
approach we investigate different behaviors in a “concurrent” setting to decide
how to behave in different situations utilizing the learned strategy.
3T acSimu on
The trac simulation used here is based on an underlying geographic information
system (GIS), built on the toolkit
GeoTools
1
and implemented in JAVA. The
simulation system is designed to simulate urban trac scenarios with different
kinds of road user, e.g., cars, trucks and bicycles on up-to-date cartographical
material. Tra
c rules like the way of right are implemented, giving the car
coming from the right side the higher priority if the corresponding road is equal
righted as the other road. In this section we provide a brief summary of the
simulation system. A more detailed description can be found in [5]. The use of
this particular simulation system is not mandatory for this approach. However,
having available the source code allows for direct integration of control behavior
and coupling of the machine learning component.
The road map is modeled by means of a graph datastructure. Each road
is represented by a data structure
EdgeInformation
(EI). Multiple EIs can be
connected with help of a
NodeInformation
(NI). EIs store information about
the corresponding road (e.g. attributes like “roadname”, “maximum velocity”,
“number of lanes”, “priority for the right of way”, ...) and administrate the
simulated road users. Each NI stores information about the connected EIs like,
e.g. rotation directions between each two connected EIs, etc.
The behavior of the simulated road users is based on the well known Nagel-
Schreckenberg model (NaSch) [13]. The NaSch model partitions a road in cells
with length of 7.5m, what is not sucient for urban scenarios. Therefore, we
adapted the model by removing the cells and enabling simulated road users to
1
http://www.geotools.org
