Supplementary material for the paper:
From Solitary to Collective Behaviours:
Decision Making and Cooperation
Vito Trianni, Christos Ampatzis, Anders Lyhne
Christensen,
Elio Tuci, Marco Dorigo, and Stefano Nolfi
This page contains support material of the paper "From Solitary to
Collective Behaviours: Decision Making and Cooperation" submitted
to the 9th European
Conference on Artificial Life (ECAL 2007). In particular, we
provide a detailed description of the experimental setup, which was
left out due to space limit. We also provide an extensive behavioural analysis of the
obtained results, supported by videos of the evolved
behaviours. Finally, we describe the behaviours produced by all
controllers that make use of communication.
Communicative Behaviours
Communication characterises also the behaviour produced by some
controllers that do not belong to class C: C4,C7,C16,C19. The signalling behaviour and the modalities with which perceived
signals are exploited significantly differ among these controllers,
and will be analysed individually.
top
Controller C4
Controller C4 belongs to class U, but it presents an interesting behaviour
that exploits communication. A continuous signal is emitted when s-bots
are positioned over a white floor, while no signalling is performed
over the circular band. Whenever a signal is perceived, s-bots either
move straight or they loop over the circular band. S-bots leave the
circular band as soon as no signal is perceived, and they move
towards the opposite side. Finally, s-bots avoid each other by changing
the direction of motion.
In environment B, the above rules produce an oscillatory
behaviour of the s-bots, which travel back and forth from the circular
band. However, s-bots never remain close to each other for a long time,
and generally obtain low fitness.
The communication strategy
employed here has the side effect of synchronising the movements of
the s-bot, in a similar way to what described in another study on
synchronisation (see Trianni and
Nolfi, 2007). In fact, s-bots leave the circular band only when no
signal is perceived. Given that signalling ceases only when all s-bots
are on the band, it results that they leave the band
synchronously.
When placed in environment A, s-bots perform the same
oscillatory movements. However, in this case s-bots have the chance to
find the way out and to move away from the arena centre. When this
happens, s-bots emit a continuous signal, which is perceived by the
other s-bots and is exploited to continue looping over the circular band
and eventually exit. Similarly to what happens with class C
controllers, communication is exploited here to share the information
belonging by the individual s-bots: if an s-bot finds the way out, it emits
a signal that is successfully exploited by the rest of the
group.
top
Controller C16
ControllerC16 belongs to class B, therefore it produces a systematic search
behaviour when s-bots are in state S, and a "bouncing" aggregation when
s-bots switch to state C. However, the switch from state S to state C is performed
exploiting sound signals, as described below.
The robots do not signal while in state S. However,
the first s-bot that switches from state S to state C emits a signal, which
triggers the behavioural switch also in the other s-bots (see the video:
s-bots shortly signal before leaving the circular band). In this way,
the aggregation process is more precise and faster, and results in a
high performance of the group.
top
Controller C7
Robots controlled byC7 also make use of communication. However, in this case they exploit it
to allocate different roles within the group.
While robots are looping on the circular band in
search of the way out, they emit short sound signals in order to decide
who takes the leader role. If the way out is not encountered, eventually
one s-bot starts emitting a continuous tone and stops on the circular
band, therefore becoming the "leader". The other s-bots react to the
perception of a continuous signal by taking a "follower" role: they
continue to travel on the circular band without signalling, until they
aggregate with the leader. The main drawback of this strategy is that
the role allocation may take a long time, sometimes exceeding the
duration of a trial. Aggregation may also take a long time, because
s-bots may need to travel across the whole circular band before meeting
the leader. The variability of the role allocation process deeply
influences the success rate in environment B.
When placed in environment A, the s-bots that find the
way out stop any signalling, therefore dropping out from the role
allocation process.
top
Controller C19
ControllerC19 belongs to the class M. Apparently, the group dynamics produced by this controller are
identical to the other controllers of the same class, but they present
a much smaller coverage of the circular band, which contrasts with the
individual decision making mechanism described above. In this case, s-bots exploit communication
to collectively switch from state S to state C, as described below
s-bots integrate over time the intermittent sound
signals they produce while searching the circular band for the
way out. The more s-bots are contemporary searching, the more signals are
produced and integrated over time, the sooner s-bots switch from state S to
state C. In order to confirm this hypothesis, we performed some test
varying the group size from 1 to 4, and we observed that larger groups
are faster in taking a decision than smaller ones.
In environment A, robots stop signalling when they find
the way out, and therefore they slow down the integration over time
performed by those robots that are still searching, which remain in
the state S for a longer time. This communicative behaviour is very
interesting not only because it exploits the signals emitted by the
group to take a collective decision, but also because it takes into
account the amount of signals perceived and the time at which those
signals are emitted.
top
