IEEE Computational Intelligence Magazine - November 2021 - 64

TABLE II Performance Comparison Among Bi-MeMAS and Existing Approaches after 2000 Learning Trials. (SR: Success Rate, MN:
Memotype Number, IN: Imitation Number, SN: Ratio of Increased SR to Increased MN, SI: Ratio of Increased SR to IN, RT: Running
Time. " á " , " + " , or " - " denote that the method is statistically similar, better, or worse than Bi-MeMAS.)
SR
Metrics
BASELINE
eTL
MeMAS
FORWARD IMITATION
BACKWARD IMITATION
Bi-MeMAS
Mean
60.4 -
70.3 .
65.2 -
63.8 -
70.6 .
71.2
Std.
4.13E-03
7.42E-03
8.76E-03
1.18E-02
8.96E-03
6.22E-03
Mean
144 +
183 -
176 -
163 +
176 .
172
MN
Std.
5.93
5.37
6.81
8.33
4.61
4.55
SR of 71.2% after 2000 learning trials. In particular, a significant
drop of 6% in SR under MeMAS as compared to BiMeMAS
was observed at the end of learning. MeMAS, which
considers an elitism-based imitation strategy, tends to blindly
believe in the knowledge transferred from the agent with the
best SR. As the exploration experience of each agent is unique,
the transmitted knowledge can be misleading and unsuitable
70
Mean
-
6.12E+04 -
1.11E+05 -
9.47E+04 -
3.89E+04 -
1.34E+04
IN
Std.
-
9.44E+03
1.545E+04
1.134E+04
6.91E+03
3.94E+03
SN
Mean
-
0.255
0.150
0.175
0.319
0.376
SI
Mean
-
1.62E-04
4.33E-05
3.59E-05
2.62E-04
8.06E-04
RT
Mean
9.23
29.62
24.76
26.75
14.79
14.28
for solving certain unseen circumstances. These results verified
the efficacy of Bi-MeMAS in reducing the impact of blind reliance
during social interactions in the learning process.
Although eTL exhibited competitive learning performance
50
30
Baseline
MeMAS
eTL
Bi-MeMAS
10
500
1,000
Learning Trails
FIGURE 8 Success rate of agents under Baseline, MeMAS, eTL, and
Bi-MeMAS in completing MNT missions.
190
×104
15
130
10
70
Baseline
MeMAS
eTL
Bi-MeMAS
10
500
1,000
Learning Trails
FIGURE 9 Memotype number of agents under Baseline, MeMAS, eTL,
and Bi-MeMAS in completing MNT missions.
1,500
2,000
MeMAS eTL Bi-MeMAS
FIGURE 10 Imitation number of agents under MeMAS, eTL, and
Bi-MeMAS after completing 2000 MNT missions.
1,500
2,000
in SR, Bi-MeMAS achieved a greater level of knowledge generalization
capability and knowledge transfer efficiency. It can
be seen from Table II and Figures 9 and 10 that the agents
under Bi-MeMAS generated 172 memotypes (MN) at a cost of
1.34E+04 times of imitation. This was much less than eTL,
which generated 183 memotypes but incurred a significantly
higher cost of 6.12E+04 in terms of IN. Moreover, the proposed
method reported an average RT of 14.28 s for completing
2000 learning trials, which was much faster than MeMAS
and eTL at RT costs of 24.76 s and 29.62 s, respectively.
To facilitate a clear understanding, SN and SI were defined
to quantitatively estimate the efficacy of social imitation in
increasing the SR regarding knowledge generalization and imitation
cost. Notably, SN and SI denote the ratios of the
increased SR over the increased MN and imitation numbers IN,
respectively, compared to the Baseline method. Typically, larger
SN and SI values are preferred, as this implies higher knowledge
generalization performance and higher imitation efficiency. The
proposed Bi-MeMAS obtained the highest value of SN, i.e.,
0.376, among all methods, demonstrating the effectiveness of
the generated memotype knowledge. Furthermore, Bi-MeMAS
also reported the highest SI value of 8.06E-04, which was
5
64 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | NOVEMBER 2021
Memotype Number
Success Rate
Imitation Number
IEEE Computational Intelligence Magazine - November 2021

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - November 2021
