Information Technology Reference
In-Depth Information
(a)
Grid topology.
(b)
An example of a learned wake-up schedule for duty cycle
of 5%.
(c)
Average end-to-end latency for different duty cycles.
Fig. 5.
Experimental results for the grid topology
still slightly shifted in time). At the same time nodes on the same hop have learned to
desynchronize their active times similar to the mesh topology.
The result of applying our learning approach in a grid topology for various duty cy-
cles can be observed in Figure 5(c). It displays the average end-to-end latency of the
network when using synchronicity and (de)synchronicity respectively. Here again error
bars signify one standard deviation across
30
runs. Due to the high data rate, when us-
ing S-MAC nodes are incapable of delivering all packets for duty cycles lower than
2%
.
This reduced performance at low duty cycles is due to the large number of collisions
and re-transmissions necessary when all nodes wake up at the same time. The learning
approach on the other hand drives nodes to coordinate their wake-up cycles and shift
them in time, such that nodes at neighboring coalitions desynchronize their awake pe-
riods. In doing so, nodes effectively avoid collisions and overhearing, leading to lower
end-to-end latency. When nodes coordinate their actions, they effectively reduce com-
munication interference with neighboring nodes. This behavior results in lower amount
of overheard packets, less collisions and therefore fewer retries to forward a message,
as compared to the fully synchronized network. Nevertheless, at very low duty cycles
the active time of nodes is too short to forward all messages and therefore, similar to
the line topology, the network experiences traffic congestion.
3.3
Discussion
We would like to discuss here the convergence time of the learning agents. The implicit
exploration scheme, described in subsection 2.5 makes nodes select different actions in
the beginning of the simulation in order to determine their quality. As time progresses,
