IEEE Computational Intelligence Magazine - November 2021 - 68

^h and eTL^h This result verifies
IN 150 000.
,
IN ,.50 000.
the efficacy of the proposed bidirectional imitation strategy in d
-MeMBi, particularly in deciding when and how an agent shall
identify and imitate a more suitable expert agent, hence leading
to the reduction of misleading and redundant knowledge transfer
in social interaction.
In summary, the proposed d-MeMBi with attention control
and bidirectional imitation strategies is superior because it
TABLE III Performance of Different Methods with Different
Numbers of Agents in UT2004 (KD: Kill Death Ratio, MN:
Memotype Number).
KD
Metrics
4V4 BASELINE
4V4 MeMAS
4V4 eTL
4V4 d-MeMBI
2V2 BASELINE
2V2 d-MeMBI
6V6 BASELINE
6V6 d-MeMBI
Mean
1.194
1.327
1.583
1.629
1.309
1.687
1.258
1.472
Std.
0.231
0.189
0.219
0.150
0.221
0.209
0.289
0.247
Mean
118
115
110
102
64
50
158
146
MN
Std.
4.291
3.844
4.056
3.864
3.539
3.181
5.306
5.193
generates more memotype knowledge inside the agent's mind
universe, leading to the improved learning performance (i.e.,
combat capability) of agents while reducing the imitation
cost significantly.
0.4
0.8
1.2
1.6
2
02 46 0 24 6
Baseline
Baseline
δ-MeMBi
Learning Steps
(a)
×104
0.3
0.7
1.1
1.5
δ-MeMBi
Learning Steps
(b)
×104
2) Study on the Number of Agents
The 2v2 and 6v6 Deathmatch game scenarios were used to test
the generality of the proposed d-MeMBi. Note that all experimental
settings except the number of agents were kept consistent
with the aforementioned experiments on the 4v4
Deathmatch scenario. The complete results of agents pertaining
to KD and MN of d-MeMBi and Baseline are summarized in
Table III. Their corresponding learning trends are depicted in
Figures 19 and 20, respectively.
Notably, in the 2v2 scenario, the agents under d-MeMBi
obtained a significantly better performance with a KD value of
1.68, which was 0.378 higher than the Baseline method. With
the increase of agent numbers, the learning performance of
agents under d-MeMBi deteriorated slightly as the complexity
of the environment increased and the behaviors of multiple
agents became more sophisticated. It took more time for each
agent to explore the environment and generate useful knowledge
that could be imitated by others. According to the results
in Table III and Figure 19, agents in the 6v6 scenario under d
-MeMBi achieved a KD value of 1.472,
which was about 0.214 lower than the
KD performance of the agents under d
-MeMBi within the 2v2 scenario.
Nevertheless, in all three learning
scenarios, the proposed d-MeMBi reported
a consistently higher KD performance
at a cost of much lower MN
than Baseline. The results thus verify the
generality of d-MeMBi for enhancing
the learning performance of memetic
agents in completing UT missions by
generating more efficient and effective
memotype knowledge.
FIGURE 19 Kill Death ratio of agents under Baseline and d-MeMBi when completing learning
missions in Unreal Tournament 2004. (a) 2v2; (b) 6v6.
VI. Conclusion
25
40
55
70
Baseline
δ-MeMBi
Learning Steps
(a)
×104
110
130
150
90
02 46 02 46
Learning Steps
×104
(b)
FIGURE 20 Memotype Number of agents under Baseline and d-MeMBi when completing learning
missions in Unreal Tournament 2004. (a) 2v2; (b) 6v6.
Baseline
δ-MeMBi
This study investigated the use of memetic
multi-agent systems for developing
human-like intelligent agents based on
the essential backbone of a memetic
automaton. To enhance the efficacy of
the conventional MeMAS for more
effective problem solving, meme-inspired
designs were enriched with meme internal
and external evolution.
First, a memotype expression process in
meme internal evolution was proposed
with an intensity attenuation parameter
to control agents' perception of the
value of information captured from different
sensors. To adaptively adjust the
68 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | NOVEMBER 2021
Memotype Number
Kill Death Ratio
Memotype Number
Kill Death Ratio
IEEE Computational Intelligence Magazine - November 2021

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