IEEE Computational Intelligence Magazine - November 2021 - 63

First, for the purpose of comparison, the memetic agent
without attention intensity control is defined as the conventional
MeMAS (hereafter labeled as MeMAS). Furthermore,
according to empirical tests by grid search, the memetic agent
with the intensity attenuation parameters
d nemi 08 .
=
d
best-MeMA ).S
and
dtarget 02 .= had the best performance throughout the learning
process. Therefore, this setting was defined as the upper bound
in the study (namely,
To verify the efficacy of the proposed MeMAS in identibest-MeMA
,S and basic MeMAS
dfying
effective d for improving the learning performance, the
performance of
d-MeMA ,S d
was compared. Figure 6 depicts the learning trends of the
agent in SR under MeMAS,
d
d
best-MeMA ,S and the proposed
dd-MeMA
.S The SR of the proposed MeMAS is lower than
best- MeMAS and almost the same as the basic MeMAS in the
early learning stage. This is because the fitness (i.e., SR) of the
memetic agent is low at the start of the learning process, and
the value of d is adjusted very slowly according to the definition
in Equation 11. When the learning process progresses,
d- MeMAS tends to have a similar performance in SR as
best-MeMA ,S
d
which is consistently higher than the basic
MeMAS. This result highlights the efficacy of the proposed
d- MeMAS in identifying an effective d for completing the
MNT tasks.
Furthermore, Figure 7 depicts the learning curves of the
memetic agent in MN under MeMAS,
proposed d-MeMA .S The memetic agent with d
d
best-MeMA ,S and the
best- MeMAS had
the lowest number of memotypes (MN), indicating that an
effective d is capable of improving the agent's learning performance
by generating more effective memotype knowledge.
During the learning process, MeMAS
dgenerated
fewer
memotypes than the basic MeMAS yet still obtained consistently
higher SR throughout the learning process. This result implies
that the proposed method can enhance the learning performance
of the agent by improving its memotype knowledge
generalization capability.
2) Results of Bidirectional Imitation
The aim of the following set of experiments was to investigate
the efficiency and effectiveness of the proposed bidirectional
imitation strategy in MeMAS for completing MNT tasks. Specifically,
a multi-agent learning scenario involving six agents in
a 16 × 16 minefield was considered.
For the purpose of comparison, the proposed method,
labeled as Bi-MeMAS, was compared against two existing imitation-based
knowledge transfer methods, including the basic
MeMAS [10] and eTL [11], which were well designed in previous
studies. In MeMAS, agents always learn from other
agents with the best fitness (i.e., the highest SR) by imitating
their sociotypes, following an " imitate-from-elitist " principle.
In contrast, eTL considers a fusion of " imitate-from-elitist "
and " like-attracts-like " principles, so agents learn from other
agents that not only have a higher SR but also bear certain
similarity under a given circumstance. Meanwhile, a traditional
multi-agent learning system, wherein the agents have no
40
500
1,000
Learning Trails
FIGURE 6 Success rate of agents under MeMAS, dbest-MeMAS, and
the proposed d-MeMAS in completing MNT missions.
1,500
2,000
100
knowledge transfer capability, was considered as the Baseline
for comparison.
The complete results of agents pertaining to SR, MN, IN,
SN, SI, and RT under Baseline, MeMAS, eTL, and the proposed
Bi-MeMAS after 2000 learning trials are summarized in Table II.
Figures 8-10 depict their corresponding learning trends.
It can be seen from Figure 8 that all multi-agent learning
methods with behavior imitation achieved higher performance
in terms of SR compared to the Baseline method. After 2000
learning trials, the agents under Baseline reported a SR of merely
around 60.4%, which was significantly lower than that of
MeMAS (65.2%), eTL (70.3%), and Bi-MeMAS (71.2%). This is
simply because the social interaction in the multi-agent learning
method with imitation enables agents to benefit from the
effective sociotype knowledge transferred from experts, thus
speeding up the agent's learning in a complex environment.
Among all multi-agent (imitation) learning methods, the
proposed Bi-MeMAS achieved consistently better performance
throughout the whole learning process, reporting the highest
80
60
MeMAS
δbest-MeMAS
δ-MeMAS
100
200
300
400
500
600
700
800
MeMAS
δbest-MeMAS
δ-MeMAS
500
1,000
Learning Steps
FIGURE 7 Memotype number of agents under the MeMAS, dbestMeMAS,
and d-MeMAS in completing MNT missions.
1,500
2,000
NOVEMBER 2021 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 63
Memotype Number
Success Rate
IEEE Computational Intelligence Magazine - November 2021

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