Information Technology Reference
In-Depth Information
Once evaluated the convergence dynamics, we conclude that the considered
coordination problem can be solved by reinforcement learning techniques with
a sucient success rate for the cases of lexicons of 2 meanings and 2 signals and
that the RL method is particularly ecient in the case of teams of 5 robots.
As the lexicon complexity or the team size increases the rate of success of
convergence to a
Saussurean
communication system decreases due to the fast
growth of the problem's search space. Thus, for 5 robots teams with lexicons of
6 meanings and 6 symbols or higher the cases of complete convergence are more
scarce, as it also happens with experiments concerning teams of 10 robots and
lexicons from 4
×
4 on. In Fig. 3 the percentage of success for several such cases
are presented.
In all these experiments the learning algorithm applied is the ACO. It punishes
the association signal-meaning used in both sender and receiver when communi-
cation fails. In the case of success, the applied association is reinforced in both
communicators and lateral inhibition is also applied in the following manner: the
sender's competitive signals and the receiver competitive meanings are punished.
Population successes vs. round. 5 robots 2x2
Population suc
c
esses vs. round. 5 robots
3
x 3
Maximum succcess
30
25
20
15
10
5
0
0
10
20
30
40
50
Round
(a) 2
×
2 lexicon (MS = 20)
(b) 3
×
3 lexicon (MS = 30)
Population successes
v
s. round. 5
robots 4x4
Population suc
c
esses vs. rou
n
d. 5 robots 5
x
5
Maximum succcess
Maximum succcess
40
50
40
30
30
20
20
10
10
0
0
0
5
10
15
20
25
30
0
5
10
15
20
25
Round
Round
(c) 4
×
4 lexicon (MS = 40)
(d) 5
×
5 lexicon (MS = 50)
Fig. 4.
Convergence to a
Saussurean
communication system with different lexicon sizes
for the case of 5 robots teams. In these graphs the number of round is displayed on the
horizontal axis and the vertical axis is for the number of success. Maximum success
(MS) is indicated in each graph.
Search WWH ::
Custom Search