Information Technology Reference
In-Depth Information
Fig. 1.
Pattern learning
capturing of simulation outputs, a learning phase is initiated which leads to a
set of patterns which are extracted from simulation. Having identified these
patterns, it is possible to use the generated information for an adapted behavior
in the simulation or for the decision which simulation runs should be performed
next (i.e., what parameter sets to use in upcoming experiments). This general
pattern learning approach is not restricted to strategy learning but could also
be used in order to identify patterns about future events which are likely to be
happen if a certain situation is present. However, in this work we focus on action
selection for a current situation.
The goal of learning is to identify a strategy how to behave in certain situ-
ations. A
strategy
is defined by a function mapping a current situation to an
action to be performed:
strategy
:
dom
(
F
1
)
A
.
In order to find out which action might be the best to encounter a certain
situation, we apply a “concurrent” execution from identical simulation situations
applying all potential actions. This procedure is illustrated in Figure 2. At a
certain time step
t
n
in simulation, a situation description is computed resulting
in the
j
attributes' values (
f
1
,...,f
j
). From this time step on, the different
actions (elements of the set
A
) are applied and executed in different simulation
runs. All these runs start from the same situation at
t
n
. After some time at
time step
t
n
+
k
, the situation is assessed for the different runs and the action
a
best
leading to the best situation is stored in combination with the situation
description: (
f
1
,...,f
j
,a
best
). The training input for learning a classifier is a
set of such situation descriptions.
×
...
×
dom
(
F
n
)
−→
4.3 Utilization of Learned Strategies
In order to dynamically adapt the behavior to the current situation, the same fea-
ture computation procedure as for generating the training data is used. When-
ever a dynamic strategy decision should be performed, a tuple (
f
1
,...,f
j
)with
