Information Technology Reference
In-Depth Information
between more distant distance relations. For the implementation of this principle,
we work with the following parameters:
Distance groups:
The (source, target) pairs are collected into groups of “sim-
ilar” distance. The distances are obtained using the static values computed
at initialization.
Launch interval:
This is the time interval between two single ants generated:
The smaller the interval, the more ants are generated. In our implementation
this interval is applied to the group of most distant pairs only.
Launch delay:
This is the delay added to the launch interval for pairs of closer
distance: The closer the distance, the greater the delay.
4 Tests and Results
Our test consisted of two parts: First, we did some experiments with parameters
in order to adjust the scalability. Second, we evaluated the quality of AntScout
from a user's point of view: How well and how fast does the ant colony react to
a sudden change introduced by an operator.
The hardware on which we tested our software included an Intel Core i5-
2520M processor with 4 x 2.5 GHz and a memory of 7.6 GB applied on a 64-bit
system.
4.1 Adjusting the Scalability
In this investigation, we tested different values for launch interval and launch
delay and tested their effect to quality which was measured by the number of
ants who really reached their target and, thus, updated the pheromones. We
compared this with the scalability of the system which was measured by the
frequency in which a full garbage collection was started.
In a Java Virtual Machine, a full garbage collection is executed automatically
when there are too many unreferenced objects. Such an event is dependent on
the number of objects generated which in our case depends on the size of the
network. The effect of a full garbage collection is an intermediate halt of the
overall simulation. This has direct influence on the performance. This is why a
full garbage collection is a good measure for scalability.
As an example, Fig. 2 shows the result for the map shown above consisting
of 104 nodes.
Considering the dashed graph, it is a little surprising that the quality is not
highest with the lowest launch interval. This effect is even more evident in larger
maps. Our explanation is that a too frequent launch of ants is too fast for
processing them, and this implies that quite a few of the generated ants will not
reach their goal. On the other hand, the behaviour for larger launch intervals is
expected because it is clear that fewer ants will reach their goal if fewer ants are
generated.
Considering the solid graph, it is clear that a higher frequency of ant launch
(corresponding to smaller launch intervals) will result in more garbage collec-
tion. This explains the monotonous decline of the corresponding graph. Since a
