Information Technology Reference
In-Depth Information
was braking harshly. Huge differences in speed lead to a higher probability of
accidents and therefore a slower forthcoming because of a lane blockage where
the accident took place.
We parallelized the simulation in order to run the evaluation for our scenarios.
In our implementation, a server controls a set of simulation clients.
4
Learning Dynamic Adaptation Strategies
In this section, we describe the approach to learning dynamic adaptation strate-
gies. The underlying goal is to identify patterns from experimental results of
simulation runs and to utilize this information to dynamically adapt a system's
behavior. The following sections address the used representation for situations
and actions, the learning process, and the utilization of a learned strategy. As
the intention is to develop a generic approach that can be used in different sim-
ulations and settings, the description is on a rather abstract level in this section.
4.1 Representation
In this work, we have chosen a rather simple, straight-forward representation for
situations and actions. The selected representation allows for a direct application
of supervised propositional learning approaches like decision tree and decision
rule learning. Situations are described by a set of attributes whose values for
an actual situation are extracted from the simulation system. Each attribute
can be either numeric or symbolic. Domains - i.e., possible values for a certain
attribute - of symbolic attributes are defined by a set of different symbolic
values. The domain of a symbolic attribute
F
i,symb
is defined as
dom
(
F
i,symb
)=
{
. Domains of numeric attributes
F
i,cont
are defined by an interval:
dom
(
F
i,cont
)=[
v
l,i
,v
u,i
]with
v
l,i
,v
u,i
∈
v
i,
1
,...,v
i,n
}
.
A list of attributes (
F
1
,...,F
n
) is used for the representation of a situation. A
specific situation is represented by a list of corresponding values of the attributes
(
f
1
,...,f
n
)with
f
i
∈
n
. Potential actions (behaviors)
to be performed in a specific situation are represented by a set of identifiers
A
=
dom
(
F
i
)for1
≤
i
≤
, e.g., different speed limits that can be imposed on a road or
the decision if a variable-message sign is turned on or off.
{
a
1
,...,a
m
}
4.2 Learning Strategies
Figure 1 illustrates the principal pattern learning process. The simulations we
are addressing in our work consist of a set of parameters (e.g., number of cars in
the simulation and fraction of trucks) and underly certain random effects which
effect the simulation. In dependence of the seed value of the random number
generator, different simulation results might be generated even if identical pa-
rameter settings are used (cf. aforementioned probabilistic behaviors in Sect.
3, e.g., dallying). Thus, multiple runs of identical parameters lead to different
results (in the general case). After performance of a set of simulation runs and
