IEEE Computational Intelligence Magazine - November 2021 - 55

placed on the design of meme selection for more effective knowledge
transmission across the population. To this end, we introduce
a bidirectional imitation strategy based on agents'
estimation of the importance and/or uncertainty of decision
making in a dynamic environment. Experiments on a minefield
navigation simulation as well as a commercial video game
demonstrate the superior performance of our proposed method
compared to state-of-the-art methods.
I. Introduction
M
emetic computation (MC) is a novel computational
paradigm that incorporates the notion of
" memes " as building blocks of knowledge for
boosting the search performance of artificial evolutionary
systems [1]. In the last few decades, the advantage
of MC has been well established as an extension of the classical
evolutionary algorithms, taking the form of hybrid, adaptive
hybrid, or memetic algorithms, where a meme is
perceived as a form of individual learning procedures or local
search operators in population-based search algorithms. In
particular, most existing works have been manually crafted
for problem solving in very specific domains [2]. Nevertheless,
falling back on the fundamental definition of a meme by
Dawkins and Blackmore (i.e., as the fundamental building
blocks of culture evolution), the potential merits of memes
remain to be fully exploited. Current research of MC has
begun to explore the development of a meme-centric
" memetic automaton " [3].
A memetic automaton is a software agent/optimizer that,
emulating humans, autonomously acquires an increasing level
of capability and intelligence through embedded memes
learned independently from past experiences or through interactions
with others. In this context, the notion of a meme is
liberated from the narrow scope of a local search/individual
learning operator. Instead, it embodies potentially diverse
forms of problem-solving knowledge [4]. For example, as the
natural building blocks of meaningful information, the recurring
real-world patterns or structures of a problem domain
(e.g., trees, graphs, or artificial neural networks) can be viewed
as forms of meme representation [5], [6] that affect agents'
perceptions, beliefs, and actions. In addition, other learning
techniques, such as reinforcement learning, can facilitate learning
pertaining to memes [7]-[9].
The conceptualization of memetic automaton has inspired a
large number of meme-inspired designs or algorithmic frameworks
that form a cornerstone of memetic computation as
tools for effective problem-solving. Particularly, recent studies
have proposed a memetic multi-agent system (MeMAS) for
developing autonomous interacting agents capable of solving
complex multi-agent planning and decision-making problems
[10]. In MeMAS, memory and learning have been the main concerns
relevant to the behaviors of memetic agents. Memes are
coined as the building blocks of memory items inside an
agent's mind universe and can evolve through self (reinforcement)
learning or social interaction with other memetic agents.
In complex learning scenarios, memetic agents are often
expected to capture environmental information with multiple
sensors as sensory inputs, much like the human brain. Previous
studies have worked with large amounts of information with
the same level of sensory involvement [11], [12]. However, due
to the physiological limitation to human ability, it is impossible
for humans to perfect the processing of all stimuli at once [13],
[14]. For example, when a person is absorbed in reading, their
central nervous system mainly captures vision information and
often fails to pay attention to other information, such as loud
noises nearby. In other words, people always choose the important
stimuli to pay the most attention to, and the impacts of
other stimuli are attenuated or weakened. Hence, the core issue
of attention is the selective analysis of information. In this study,
an intensity attenuation control method is proposed, taking
inspiration from the psychologically-inspired Broadbent-Treisman
Attenuation Model of environmental attention [15], [16].
By adjusting the intensity of environmental information from
diverse sensory units, memetic agents are capable of catching
more important information that is expected to be remembered
more clearly and carefully, thus leading to more effective
knowledge generalization.
Notably, memes in a memetic automaton are natural building
blocks of meaningful knowledge and support transfer/reuse
across problems [17]-[20]. This capacity to draw on memes
from past instances of problem-solving sessions thus allows the
search to be more intelligent, leading to solutions that can be
attained more efficiently in unseen circumstances of increasing
complexity and dynamic in nature. In MeMAS, as the learning
process progresses, the beneficial meme knowledge of individuals
will be proliferated in a population during social interaction
primarily driven by imitation. Following a commonly used
" imitate-from-elitist " principle [10], MeMAS has used social
learning strategies where an agent with poor performance
always learns from an elitist with the best overall fitness performance.
However, this learning strategy may suffer from blind
reliance [21], which could hinder the learning process.
Therefore, this study goes a step forward by investigating an
interactive imitation strategy capable of determining when and
how an agent could benefit from the useful memes of others.
Note that the performance of an elite shall not be the only factor
that drives imitation. The teacher and student agents will
jointly identify imitation opportunities. Under a given circumstance,
the student agent determines whether to ask for attention
from other agents, and the teacher agent, if asked, makes a
choice on whether it is confident to provide instructions for
the imitation.
The contributions of this paper can be summarized as follows:
❏ An attention control model is introduced into the learning
process of memetic agents when the environmental information
is received from diverse sensory inputs. By adjusting
the intensity of environmental information, the proposed
method enhances agents' perception of the dynamics and
uncertainty of the environment, therefore leading to more
efficient meme knowledge generation.
NOVEMBER 2021 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 55
IEEE Computational Intelligence Magazine - November 2021

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