IEEE Consumer Electronics Magazine - March/April 2020 - 54

Unmanned Aerial Vehicles

Figure 1. Self-organizing UAV array deployed
above the city (parts are designed by freepik).

State-of-the-art swarm intelligence perhaps can
find solutions. According to Wikipedia, swarm
intelligence studies "the collective behavior of
decentralized, self-organizing systems. . .." In
other words, rather than concentrating on an
individual entity, swarm intelligence discusses a
population of entities. Relevant topics range from
population movement and crowd simulation to
cluster formation.12 As geospatial allocation for
multiple drones is similar to population movement in swarm intelligence, one of the related
approaches-Self-organizing maps (SOMs)-
could shed some light. The following section
describes in detail how SOMs work on drones.

TACTICAL GEOSPATIAL
DEPLOYMENT BASED ON
SELF-ORGANIZING
SWARM INTELLIGENCE
The SOM was proposed by Teuvo Kohonen.
The underlying model is an artificial neural network based on machine learning theories. Such
a network is composed of only two layers of neurons. Figure 2 shows the structure of a typical
SOM. The first layer, or the bottom part of
Figure 2, consists of perceptrons that are used
for receiving/perceiving stimuli from outside.
When stimuli arrive, they are processed by perceptrons. Those processed signals subsequently
affect the second layer through the interconnections between two layers, and the topology of
neurons in the second layer responds to input
signals. In SOMs, the second layer is used to
visualize changes in the topology, incurred by
outside stimuli. This layer is the front end,

Figure 2. Structure of a SOM, where the top part is
the map space, and the bottom is the input layer.

exposed to human readers, so that it is also
called "visual map space." This is why SOMs are
frequently regarded as a visualization tool due
to the capability of mapping high-dimensional
input data (e.g., stimuli) into a low-dimensional
intelligible topology. Morphing of the topology
in the visual map space is simply based on the
law of attraction. This is a critical feature that
allows a UAV array to perform autonomous
deployment. The intuition behind topological
morphing is that when a stimulus is input, the
input is fed into all the perceptrons. After the
processing by the perceptrons, only one neuron
in the map space is activated. This activated
neuron attracts its neighbors, and the proximal
neighbors move closer to the activated one. The
process of SOMs is similar to a magnet versus an
elastic net full of steel beads. When the magnet
is placed at a selected spot on the net, the magnet attracts steel beads around it. How is this
spot selected, or equivalently how is this neuron
chosen? It is based on similarity between the
stimulus and the neuron. It is reasonable for a
stimulus to be represented as a numerical vector. The question is how to denote a neuron on
the grid net as a numerical vector, so that similarity between the stimulus and the neuron can
be computed. As mentioned earlier, the two
layers of a SOM are interconnected by edges,
and each of the edges has a weight. These
weights can be viewed as those for neurons on
the grid net. In brief, SOMs work as follows. A
stimulus triggers an activated neuron, and this
neuron assimilates neighbors. This is the intrinsic mechanism of SOM morphing.
To enable autonomous deployment for a UAV
array, the array system follows the topological

IEEE Consumer Electronics Magazine

IEEE Consumer Electronics Magazine - March/April 2020

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - March/April 2020