geneura-papers / 2016-hpca-hpmoon Goto Github PK

Could you please add a sentence about the setting of the reference points for IGD calculation?

4. Same levels are considered for parameters in order to implement all four approaches. While maybe some methods with other levels of parameter could provide better results. For example, maybe the variable to variable crossover instead of the SBX crossover technique provides better results in some approaches. The parameters of each method should calibrate and then the approaches should be compared. Why were same levels of parameters considered for all for approaches?

• I do not understand the sentence “This can be explained because added migration latency….”. I understand that both the previous model and the new one use synchronous migration and, therefore, I would say that this migration is done at the same steps of the execution. It should be better explained.

8- Comparing the solution sets of the proposed methods in a visual way is highly recommended. For example, using a Parallel Coordinates plot to see the convergence and the diversity of the solution set. A reference that may be helpful is listed below:
M. Li, L. Zhen, X. Yao, How to Read Many-Objective Solution Sets in Parallel Coordinates. IEEE Computational Intelligence Magazine, 12 (4): 88-100, 2017.

(Note: it was done in previous version :/)

Reviewer 2:
If possible, could you please visually show the effect of the number of islands on the obtained solution set?
Reviewer 3:
5) The result section is weak. The behaviour of the methods in different levels of the number of islands and the length of the chromosome should discuss in more details and may be with some figures. No clear conclusion based on that which method outperforms the other ones is presented.

3. Why did the authors name Section 3.1.3 as adaptive? The adaptive is used, where some partial authorizations are allowed in one of working parameters during the algorithm iterations.

• 3.- Why a distributed NSGA-II is chosen as the baseline? Would not be fairer to use the sequential version? Related to this, how have running times been assigned? 25/100 seconds for each process (island) independently of the number of islands? If this is the way, this is not a fair comparison, as we are using more global cpu time (more machines) as the number of islands increases. May be I have not understand this correctly.

4- The main contribution of paper should clearly state at the end of introduction.

It seems to me that the setting of the population size has a large effect on the performance comparison results.
Could you please explain the reason for this particular specification (i.e., N = 1024)?
I am asking this question because we usually do not need such a large population for approximate the Pareto front of a multi-objective problem with two or three objectives.
Could you also please add some comments on experimental results from other settings (e.g., N = 256, N = 4096)?
That is, please add some comments on the sensitivity of the reported experimental results in this paper to the specification of the population size.

The usual definition https://wiki.ece.cmu.edu/ddl/index.php/Coevolutionary_algorithms is an algorithm where fitness depends on context: cooperative or competitive. I don't think that's the case here, even if Dorronsoro says so.

There are lots of grammar and in some cases spelling mistakes.

5- The research gaps are missed in the present version of this manuscript.

(IDK what the referee wants here...)

• 2- This is a journal focused on high performance computing. I see no pseudo-codes of the algorithm, communication schemes between islands, and other detail that allow understanding the proposed algorithm. In addition, a performance analysis, speed-ups, communication-computation ratios, and so on would be welcome. Some of these aspects, for example the time required for migrating individuals is mentioned as a drawback of other approaches, but it is not discussed at all.

• An intuition of the meaning of the term "overlapping sections of individuals" should be made clear(er) since the very beginning. For instance, in both the abstract and introduction this term is used but not defined, nor informally introduced. So, basically the true content of the paper becomes clear only with Figure 2 at page 7.

• 1.- The experimental section compares different strategies: among them and against a baseline NSGA-II. However, there are many different multi-objective algorithms in the literature (for example SPEA, MOEA/D), and also different flavors of them. I think that, at least in the scope of island-based models, the authors should demonstrate that their algorithm (or even strategy) is competitive compared to the state-of-the-art algorithms. I know that performing this kind of analysis is difficult, but otherwise does it has sense to use this strategy if the whole algorithm is not competitive?.

• The conclusions section seems inappropriate for a journal paper, and needs to be deepened.

• Regarding the formula, which is supposed to be the main contribution, does it round or truncation (equation should be properly written)? How has it been calculated, has some regression model or similar technique been used? How does influence the size of the population (it is slower as the number of islands increases)? Why a fixed population is used throughout the experiments? Has it been set according to previous experiments?

Reviewer 2
Could you please explain why c should be an integer?
It seems to me that c can be a real number such as 0.5 (i.e., 50% overlap with the adjacent block) and 1.5.
Reviewer 3

1. Why did not the authors use statistical methods (e.g., Taguchi method) to determine C?
2. How did the authors design such formula for C? It is just mentioned that they have used the knowledge of the previous tests.

It seems to me that the generations between migration is also an important parameter.
Could you please explain why this parameter was specified as 5?
Could you also please add some comments on the sensitivity of the reported experimental results to the specification of this parameter?

Could you please clearly explain how to select "one individual" in a given island? Totally random, random from the non-dominated solutions?

• Some choices of this paper deserves a deeper explanation. For instance: why did you use exactly that formula (at the beginning of page 7) to calculate the c value (extra sections to overlap). Why did you use exactly those measures of performance (HV, IGD and Spread) and not others? For instance, I think it would be appropriate to also report the obtained results on the different optimization criteria.

• In addition, I miss some discussion about the relation between islands and diversity. Is this the main advantage of the proposal? Which is the migration ratio, how does this affect diversity and, hence, results?

2- The first paragraph of introduction should discuss about the study motivation and how the proposal can improve on the state of art?

Specially from ASOCO and the ones published in 2017

• Authors claim that, for 2048, the best results for all the quality indicators are obtained by using overlapping. According to Table 3, the baseline approach shows clearly better Spread values, and it not that clear the A option to perform better.

1- Abstract is too long. It is clearly below of standard publications. The abstract should be shorten. The authors should discuss about the main innovation of their study and the interesting results from the evaluations.

3- Why did not the authors use statistical methods (e.g., Taguchi method or response surface method) to determine C?

6. What is "point (1,9)" in the first line of Page 11?

• The proposed method has been tested only on one benchmark (ZDT). Although the benchmark is well known in the community, I think this is not enough to demonstrate the generality of the approach. Other benchmarks of different nature, and possibly also real-life applications should be used as test cases.

• I understand that an island-based model is designed for dealing with large problems, which cannot be executed in one computer. In the experimental section, ZDT functions are used, up to 2048 variables, and execution times up to 100 seconds. I would not say that this is a problem that requires HPC. I think that the advantages of this proposal should be tested on large artificial, or even better, real-world problems.