Академический Документы
Профессиональный Документы
Культура Документы
Review of pap196s2 by
Reviewer 1
Comments for Authors:
The authors investigate the combination of
evolution and learning strategies in the
optimization of dynamic NK landscapes. The idea of
generating dynamic optimization problems by using
NK landscapes with changing parameters is not new.
See:
Eriksson, R., & Olsson, B. (2004). On the
performance of evolutionary algorithms with life
time adaptation in dynamic fitness landscapes. In
Proceedings of the 2004 Congress on Evolutionary
Computation (Vol. 2, pp. 12931300). IEEE.
The algorithms used here are not new: they were
proposed and used in other papers [1214]. In this
way, it seems that the contribution of the paper
is an analysis of the algorithms in the dynamic NK
landscapes. In my opinion, it is not a
contribution strong enough for acceptance of the
paper in GECCO. In fact, an analysis of the
behavior of the algorithm from a practical or
theoretical point of view could be interesting.
However, the experiments presented in the paper
are very simple.
First, if I understood correctly, the number of
evaluations is different for the different
algorithms. The two algorithms that use learning
perform much more function evaluations than the
other two algorithms. In this way, it is not
possible to affirm that a better performance is
due to the use of learning or of the use of more
function evaluations. Second, the dimension of the
problem (N=20) is very low: we will have 2^20
possible solutions, what is not too much (if the
problem is the use of exhaustive search to check
the error, then the authors could use dynamic
programming and the adjacent model). More
experiments are needed.
Review of pap196s2 by
Reviewer 2
Comments for Authors:
The Baldwin effect describes improvement in an
agent’s fitness that takes place through learning
rather than genetics. This paper draws a
distinction between two forms of learning: asocial
learning, in which an agent learns through
individual experience, and social learning, in
which the agent learns by imitating other agents.
It then investigates evolutionary strategies that
combine these two learning mechanisms, in the
context of dynamic NKlandscapes.
This is an interesting investigation, based on
sound ideas, but its presentation is rendered
difficult and frustrating by the style and
language in which it is presented.
There are odd phrases and illchosen words, but
most annoying are the many unnecessary words and
phrases that clutter the paper. For example, the
last paragraph in Section 2 could be restated as:
“Most research on evolution and learning in NK
landscapes assumes fixed environments. We explore
the effect that combinations of social and asocial
learning might have on evolution in a dynamic
landscape.”
In Section 2.3, settle on one set of vocabulary
and stick with it. Also, don’t use citations as
nouns, and in citing the work of others, mention
only last names. Beware of the passive voice.
In general, be suspicious of every adjective,
adverb, and introductory phrase. Most of them can
disappear. Overall, the paper could probably be at
least an entire page shorter if every paragraph
were pared down to its essentials.
Review of pap196s2 by
Reviewer 3
Comments for Authors:
The manuscript intends (and succeeds) to
demonstrate that minimal models of individual
learning (IL) and social learning (SL), when
combined, improve the populations adaptation to a
dynamic fitness landscape. The domain is
introduced informatively, the methods are
explained with enough detail and the results
plotted are legible with easy to draw conclusions.
Individual, iterative trialanderror learning
contributes most greatly over the baseline GA, but
when social learning is included as well, there is
another marginal but consistent improvement in
adaptation. This effect is well explained in the
discussion, but probably most appropriately
related to the increase in plasticity that affects
both IL and SL in the frequency of '?' alleles
that allow bit flipping in the phenotype.
The topic is of suitable relevance to the GA track
of GECCO.
This study investigates a particularly minimal
model of both learning and imitation. For
learning, an individual performs iterations of
some trialanderror bit flipping on a designated
'?' allele, akin to mutation. For imitation, these
same '?' positions are used to copy bits from the
best individual of the previous generation. While
this is simplified to the point that it may not be
well applied to practical problems, it does serve
to illustrate the effectiveness at combining
individual learning and social learning for
adapting to dynamic fitness landscapes.
Another thing to note about the individual
learning scheme employed here is that it is
particularly expensive compared to other GA/ES as
it must have multiple fitness function evaluations
for each individual, per generation. This is not
an issue for the NK model, but in useful
applications, that usually have computationally
expensive fitness functions, would be a major
concern.
As the work intends to primarily measure the
populations adaptation to a dynamic fitness
landscape, it is understood that K is changed
randomly between 5 and 40 epochs to identify more
dramatic effects in the fitness comparison.
However, it seems to be a very artificial and
unnatural scenario with less relevance to real
world applications than a continuously varying
parameter that controls the fitness landscape.
This would relate more directly to environmental
drift in physical systems.
The submitted manuscript is wellwritten and clear
enough to be acceptable in its current form. There
are some minor issues spotted on the first page:
'...' on Para1 in Introduction.
'studies studying' > 'studies on'
'both forms' > 'either form'
Review of pap196s2 by
Reviewer 4
Comments for Authors:
The general idea of studying social learning in
the context of a dynamic problem is reasonable and
very interesting. The results also look
interesting. Unfortunately, the explanations given
in the paper are quite unclear and poorly
structured.
The “Baldwin effect” is mentioned at the beginning
as being important but is not explained.
The abstract and/or introduction should at least
give a brief summary/reminder.
The introduction also needs to be clear of how the
Baldwin effect will be used in the paper it only
really describes mechanisms for social learning,
and not any effect on the genotypes.
This should be discussed much more clearly.
The explanation of the asocial learning in section
3.2.3 is not at all clear. The role of the “?” is
not properly explained. E.g. in “new behaviour is
more adaptive” there is no clear description of
how the “adaptiveness” of a behaviour is measured.
Minor Issues
Page 1 col 2.
“has been received” to “has received”
page 2. col 1.
in the list of 3 questions, each question should
end with “?”.
sec 2.2 The starting sentence seems to repeat what
is said in introduction, but still not does
explain the Baldwin effect.
page 3 col 2
“loose” to “lose”
Review of pap196s2 by
Reviewer 5
Comments for Authors:
This paper investigates the effect of learning
strategy on evolution process when dealing with
dynamic rugged landscape. The experimental results
show that social learning performs better when the
NKlandscape is more rugged and more variable.
This paper is well organized and written.
The suggestion for the authors are as follows:
1. Please explain why mutation is not allowed in
all experiments. In my mind, mutation is a basic
operation in genetic algorithms.
2. Please point out the definition of question
mark in section 3.2.4.
3. Please indicate what the axes in figures 1 and
2 represent.
4. Please provide the specific formula of
plasticity.
Review of pap196s2 by
Reviewer 6
Comments for Authors:
This is an interesting, wellwritten paper that is
easy to follow. A have a few comments.
1) Stuart Kauffman, cf. reference [9] in the
paper, also defined coupled fitness landscapes.
These
could and maybe should also be studied. The
coupling makes them dynamic in a more natural way
then the landscapes used in the paper.
2) Wouldn't "The Baldwin Effect in Dynamic Rugged
Landscapes" be a better title than the one used?
3) Some ICGApapers that are relevant here are not
referenced, e.g. a)
Richard K. Belew: When Both Individuals and
Populations Search: Adding Simple Learning to the
Genetic Algorithm. ICGA 1989: 3441, b)
Richard K. Belew: Evolution, Learning, and
Culture: Computational Metaphors for Adaptive
Algorithms. Complex Systems 4(1) (1990), and c)
Bernard Manderick, Mark de Weger, and Piet
Spiessens: The Genetic Algorithm and the Structure
of the Fitness Landscapes. ICGA 1991: 143150,
that uses the NKlandscapes the study the
performance of the GA.
Review of pap196s2 by
Reviewer 7
Comments for Authors:
This paper investigates the effect of individual
learning and social learning on dynamic
optimization problems using a modified NK
landscape benchmark in which the K parameter
changes over time at regular intervals. The paper
is very wellwritten with a clear objective and
research hypothesis in mind. The paper confirms
the benefit of learning along an evolutionary
process and emphasizes their importance and
significance in the context of dynamic
optimization. The paper fall shorts with respect
to addressing the significance of its findings in
relation to a broader set of problems, such as
continuous problems. It would be nice if the
authors comment on how their findings can be
generalized to continuous domain.
The paper is also detached from the broader
related literature such as its connections to
individual and social learning in particle swarm
optimization and swarm intelligence as well as
evolutionary dynamic optimization. More details is
given below.
I think the subject of this study is also related
to particle swarm optimization and swarm
intelligence in general.
What is the significance of the of the current
work with respect to static environments? I have a
feeling that the notion of combining IL and SL
have broader consequences than just being useful
in a dynamic settings. I suggest that the authors
investigate the merits of their work in static
environments as well.
Now after reading the rest of the paper, I can see
that there are works on the effect of learning in
static environments.
"received a lot of attention because this form of
learning" > a bit informal. Please rephrase.
Page 2, LHS, items (i) to (iii). Are these
punctuated correctly? These are obviously
questions; do they need question marks?
It is necessary to address the body of research on
evolutionary dynamic optimization. I am not expert
in that area, but the current work seems to be
completely detached from the evolutionary dynamic
optimization literature. Please see the following:
Nguyen, T.T., Yang, S. and Branke, J., 2012.
Evolutionary dynamic optimization: A survey of the
state of the art. Swarm and Evolutionary
Computation, 6, pp.124.
Jin, Y. and Branke, J., 2005. Evolutionary
optimization in uncertain environmentsa survey.
IEEE Transactions on evolutionary computation,
9(3), pp.303317.
Branke, J., 2001. Evolutionary approaches to
dynamic optimization problemsupdated survey.
Branke, J., 1999. Evolutionary approaches to
dynamic optimization problemsa survey.
Figure 2: why is there a marked difference between
the initial average fitness of the EVO versions
with learning and those without learning at
iteration 0? The versions with learning always
start with a better fitness? Why is that? I
presume they all should start from relatively the
same average fitness. This is very important issue
and needs to be investigated and addressed by the
author(s).
There are also some performance metrics commonly
used to compare algorithms in evolutionary dynamic
optimization.
The paper lacks proper statistical significance
test (preferably a nonparametric test).
Committee Comments
& Notes
Committee Comments for Authors
Our apologies that this paper was inadvertently
assigned an unusually large number of reviewers.
We hope that the comments will be useful for the
researchers.