
Anonymous peer review system
Composition of review comments
1. Major Contribution of the paper
2. Technical Accuracy
3. Presentation
4. Adequacy of Citations
Frank, brutal, sarcastic, or even occasionally wrong criticisms
If you receive a critical comment on your manuscript, it is very normal and do not be surprised.
Examine carefully all the comments made by the reviewers and by the Associate Editor.
Differentiate those comments that must be fully addressed from those that are desirable to be addressed.
Try to fully address the former and discuss the latter within your capability.
Respond to EACH and EVERY point raised by the reviewer/referee.
If posibble, respond side by side with the original points raised.
If you must disagree with the reviewer on some points, do it professionally with no rancor or sarcasm.
A cover letter to the corresponding editor that you have responded to each of the referees in detail and that your revised paper now merits reconsideration and review.
Separate and detailed pointbypoint response to everyone of the referee's comments and show HOW YOU REVISE THE TEXT.
Long list of explanations to the reviewers but no revision on the text is not acceptable.
Never argue with reviewers and editor: our first goal is to get paper published rather than making enemies.
Know when to give up for a particular paper worthwhile for no more efforts.
Major Contribution of the paper:
The paper describes the application of Rough set theory to remannufacturing. Machine learning is used to predict a feasibility plan to remanufacture a given set of products. An iterative reinforcement learning process is applied to improve the learning process, however, only simulation results are presented. It would be beneficial if experimental results can be shown where validation of the presented approach can be made. If experimental results are not available, it is suggested that the authors point out if experiments will be conducted or what are the next research activities that are planned in the area.
Technical Accuracy:
The presented methodology is novel and has significance. It would be good to have experimental results which can be used to validate the simulation results. If these data is not available, the authors should include a description of future work and other validation methodologies that are being considered.
Presentation:
The presentation is clear and follows a logical sequence. However, the conclusions need to be improved by describing research directions comparision studies and validations techniques that are or will be used.
Adequacy of Citations:
Citations are adequate, however, it is suggested that other publications should be included where comparisons between different approaches can be made.
Major Contribution of the paper:
This paper presents a machine learning approach to determine the feasible plans of a remanufacturing systems based on rough set theory, including selection of training datasets, attribute discretization, atrribute reduction, rule generation and decision making. Iterative reinforcement is applied to improve learning performance. The numerical examples provided in the paper show that the feasible plans can be determined efficiently with high confidence level. In my opinion, the method provides an opportunity for large scale optimization problems in remanufacturing system study. However, due to the following comments, I suggest the paper could only be accepted after major changes.
I agree with the authors that the method can be applied to optimization problems in future work. However, it is not covered in this paper. The paper now only dresses the determination of feasible plans. In particular, the method is not efficient for the models presented here. The reasons are:
(1) Although model (2)(5) introduces oprimization, howerver, it is not discussed in the paper. As it is claimed by the authors, only the feasibility determination problem is considered.
(2) The remenufacturing system model ((2)(5)) is a simplified model, where many complexities, such as scheduling, routing, reentrant loops, random precessing times or unreliable repair machines, finite buffers capacities, etc., have been ignored. Moreover, it is siplified again in later discussions. For instance, the arrival rates are constant, no assumption on the order of new parts is discussed(here by we have to assume they are constant also), etc. No randomness is found in the current model. Therefore, we may guess that deriving an approximate calculation formula may not be an unfeasible anymore.
(3) It is claimed in the paper that using this method could increase the efficiency eight times compared with simulation. However, simulation for 1000 plans only take 4 minutes. In other words, each plan may take 0.24 second only. In this case, it is absolutely acceptable. In addition, to use the method, a lot of efforts have to put on selection of training datasets, attribute discretization, attribute reduction, rule generation, etc., which may take much longer time, and the method itself is not 100% accurate. From above, we have to conclude that for the model introduced in this paper, the method do not increase the efficiency to determine the feasible plans.
Therefore, I think the paper is not acceptable in current format. However, if the authors introduces not only feasibility determination, but also optimization part, I think the paper would be appropriate for publication.
Technical Accuracy:
(1) The paper spends a lot of efforts on introducing the basic ideas of rough set theory. Although it is important, especially for readers who are not familar with the theory, it indeed distracts the theme of the paper. I think this part can be siplified dramatically. In fact, many concepts introduced here are never used in later discussions. In particular, it would be better if the authors use a remanufacturing system example to replace the example of temperature, headache and fever to explain the theory.
(2) The paper discussed several difficulties in determining the feasible plans for manufacturing system on pages 3 and 4. I agree with the conclusions. However, many difficulties are not shown in the model the author considered in Figure 1. For instance, there is no reentrant loops, buffers seem to be infinite (since no capacity of buffers in shown in (2)(5)), arrival rate and remanufacturing cycle times seem to be deterministic (no distribution, or mean or variance are introduced, and it is indeed assumed to be constant in later example), etc. Thus, we have to conclude that the model studied by the authors is not difficult as they claimed to be.
(3) The authors claim that it is important to select a proper dataset for training, however, it is not mentioned how to do it. A few comments on it will provide more clear idea to the readers.
(4) The arrival rate, allocated capacities, and required probabilities are introduced in the training dataset, however, others (e.g., order for new parts, O, cycle time, Tc, requeired cycle time, d, etc.) are never introduced. Should them be included in training dataset?
(5) On page 7, it is said 'if no rule is found, the most frequent outcome in the training data is chosen', what is the reason for it? How much discrepancy it could bring into the decision making? Such questions should be answered or explained.
(6) On page 7, formula(15) uses 'Ceil' function to 'round the elements of x to be the nearest integer'. Typically, 'Ceil' is used to round up, not the closest integer. In other words, ceil(2.4) = 3, not 2. In fact, tables 5 and 6 show that they round up.
(7) How to round probabilities Pb, it is not explained. According to tables 5 and 6, we guess that if it is less than required probability, it rounds to 0, otherwise to 1.
(8) Table 3 provides steps for itrative reinforcement process. After using 1000 data set to generate 5500 rules, a new 1000 dataset is used for test and results in 95% confident ratio. Then another 10000 new reinforcement samples are selected. Does those include the 5% of the tested sample dataset?
(9) It seems to me not convincible that 5500 rules can be generated using only 1000 dataset samples, could the authors provide more insights (or remark) on what kind of rules they cold be? (Mybe some rules are so basic that can be generated even without any semple?) In addition, the reinforcement selects 10000 data samples, why don't they use 10000 at first? In addtion, how are those training or reinforcement dataset selected? Randomly? How does the selection affect your learning process? Those questions should be explained or at least the readers should be referred to other papers.
(10) The discretization includes probability Pb. If different required probability Prc is selected, then it would result in different decision making. However, it does not mentioned in the paper whether the method includes training for different Prc or not.
(11) The simulation experiment is not explained clearly. For instance, how many replications, how much simulation time, warmup time, how the initial parameters are set up, etc. (Since in the example introduced in this paper, the model is simplified too much, thoes information may not affect.)
Presentation:
The presentation is ok but can be done better, in particular,
(1) The reference format is not consistent throughout the paper. Some uses [numbers], some uses name + date, etc.
(2) On page 8, it should be PIV (not P_VI) 2.0 GHz PC.
(3) On page 3, one additional 'the' is included. It should be the required cycle timne.
Adequacy of Citations:
The citations are fine. The authors could also refer to some manufacturing system research papers which are not using queueing models.
The authors deeply appreciate the reviewers' valuable comments. The paper has been completely rewritten. Motivations and background of the remanufacturing research re presented and why machine learning method is applied to sovle the remanufacturing system is explained more clearly in the Introduction. Section Ⅱ is greatly enhanced and the problem formulation of the remanufacturing system considered in this paper is presented in much more detailed than the original version. The uncertainties and complexity involved in the remanufacturing process are shown in detail. The numerical testing is completely redone for the cases close the actual system. The response to each reviewer's comment is listed individually as follows.
Q: It would be beneficial if experimental results can be shown where validation of the presented approach can be made. If experimental results are not available, it is suggested that the authors point out if experiments will be conducted or what are the next research activities that are planned in the area.
A: The experiment has been redone and the performance of the method(include efficiency and accuracy) is described clearly in Secction Ⅴ. The background of the research is explicitly presented an validation is being conducted in the actual system of a jet engine remanufacturing factory.
Q: The presentation is clear and follows a logical sequence. However, the conclusions need to be improved by describing research directions comparison studies and validations techniques that are or will be used.
A: The conclusions are revised accordingly.
Q: Although model (2)(5) introduces oprimization, howerver, it is not discussed in the paper. As it is claimed by the authors, only the feasibility determination problem is considered.
A: The entile framework and the current method are better emplained in Introduction. This is the first phase of a constrained optimization method near optimal plans in ordinal sense. In fact, determining feasibility is of significant practical value to our industrial partner. The method developed in the paper is being integrated with ordinal optimization and results will be reported later.
Q: The remanufacturing system model ((2)(5)) is a simplified model. Therefore, we may guess that deriving an approximate calculation formula may not be an unfeasible anymore.
A: Many more details in problem fromulations are added and elaborated by the figures and equations in Section Ⅱ and fomula (2)(13). It can be seen that it is impossible to determine the feasibility analytically.
Q: It is claimed in the paper that using this method could increase the efficiency only eight times compared with simulations. In addition, to use the method, a lot of efforts have to put on selection of training datasets, attribute discretization, attribute reduction, rule generation, etc., which may take much longer time, and the method itself is not 100% accurate.
A: The problems in the original experiments are over simplified, and the RST method is implemented in Matlab but the simulation implemented in C++. To have a fair comparison, the problem formulation in this paper is close to the actual applications and the methods tested and compared in the paper are all implemented in C++.
Numerical testing is completely redone and increase of the efficiency by the RST method is very sigificant. Although, election of training datasets, attribute discretization, reduction and rule generation are time consuming, these tasks are needed to be done only once. The determination on feasibility is very efficient and suitable foe real time decision.
Q: The paper discussed several difficulties in determining the feasible plans for remanufacturing system on page 3 and 4. I agree with the conclusions. However, many difficulties are not shown in the model the author considered in Figure 1.
A: Modified accordingly as shown in Remark 3.2.
Q: The authors claim that it is important to select a proper dataset for training, however, it is not mentioned how to do it. A few comments on it will provide more clear idea to the readers.
A: Modified accordingly as explained in Section Ⅳ.4.
Q: The arrival rates, allocated capacities, and required probabilities are introduced in the training dataset, however, others(e.g. order for new parts, O, cycle time, Tc, required cycle time, d, etc.) re never introduced. Should them be included in training dataset?
A: Good comments. Actually, the order for new parts is reflected by the change of rotable inventory and the related costs. Therefore, this attribute can be excluded from the training dataset. The cycle time and required cycle time need not be included because they are hidden in the attribute "Probability of cycle time requirement"(3). The above is explained in Section Ⅳ.1.
Q: On page 7, it is said 'if no rule is found, the most frequent outcome in the training data is chosen', what is the reason for it? How much discrepancy it could bring into the decision making? Such questions should be answered or explained.
A: Clarified accordingly and some details of the general RST method is omitted. Actually, if no rule fires, it is considered as undecidable.
Q: On page 7, formula(15) uses 'Ceil' function to 'round the elements of x to be the nearest integer'. Typically, 'Ceil' is used to round up, not the closest integer. In other words, ceil(2.4) = 3, not 2. In fact, tables 5 and 6 show that they round up.
A: (lost)
Q: It seems to be not convincible that 5500 rules can be generated using only 1000 dataset samples, could the authors provide more insights(or remark) on what kind of rules they could be?
A: As explained in Section Ⅳ.5 since several rules are needed to describe each relation between reducts and decision attribute.
Q: The discretization includes probability Pb. If different required probability Prc is selected, then it would result in different decision making. However, it does not mentioned in the paper whether the method includes training for different Prc or not.
A: It is clarified in the paper that the required Prc is not included as an input since it is constant to the output: cycle time probability Pb. Once Pb is abtained, the feasibility is determined immediately by direct comparison. Training for different Prc is needed if the output is feasibility. This would increase the dimension of the decision space and computational efforts.
Q: The simulation experiment is not explained clearly. For instance, how many replications, how much simulation time, warmup time, whow the initial parameters are set up, etc.
A: Modified accordingly as shown at the end of Section Ⅱ and the start of Section Ⅴ.
Archiver手机版科学网 ( 京ICP备07017567号12 )
GMT+8, 2021923 08:37
Powered by ScienceNet.cn
Copyright © 2007 中国科学报社