Information Technology Reference
In-Depth Information
in nature and simply can not interfere with each other. This may, for
example, be the case for actions which affect the environment and
actions which represent internal reasoning processes of an agent.
Combined decomposition
Obviously, the two basic approaches presented may be combined.
Using spatial decomposition and action-dependent decomposition
in a combined fashion may circumvent the problem of splitting a
model into disjoint sets. In a first step, spatial decomposition may
be used to split a model into few (large) partitions. Secondly, the
resulting partitions may be further split up using action-dependent
decomposition.
The main benefit of combined decomposition is that it is easy and
allows fast computation. On the negative side, combined decomposi-
tion may not always exploit the maximal possible parallelism. The
total number of threads used by this approach is
g
a
g
(with
g
the
number of spatial regions and
a
g
the number of partitions resulting
from action-dependent decomposition within region
g
). If it is not
possible to compute the required number of action-independent par-
titions (e. g., due to a huge number of interfering types of actions),
some of the available threads may remain idle.
×
Self-optimizing decomposition
A further approach for partitioning a model is to employ methods and
techniques of self-optimization. The main problem of all partitioning
strategies is that the quality of a partitioning can only be evaluated
afterwards. In general, this is done by measuring the runtime and
evaluating whether the current partitioning strategy led to significant
speedup. All attempts to evaluate the partitioning in advance are of
secondary value. This is because the measurement of direct indicators
(like e.g., equality of distribution of agents to partitions, minimization
of agent interaction, minimization of network trac, etc.) which may
