Next Article in Journal
Single-Image Dehazing Based on Improved Bright Channel Prior and Dark Channel Prior
Next Article in Special Issue
Integrated Circuit Design of Fractional-Order Chaotic Systems Optimized by Metaheuristics
Previous Article in Journal
A Novel Power Distribution Strategy and Its Online Implementation for Hybrid Energy Storage Systems of Electric Vehicles
Previous Article in Special Issue
Differentiable Constraints’ Encoding for Gradient-Based Analog Integrated Circuit Placement Optimization
 
 
Article
Peer-Review Record

Procedural- and Reinforcement-Learning-Based Automation Methods for Analog Integrated Circuit Sizing in the Electrical Design Space

Electronics 2023, 12(2), 302; https://doi.org/10.3390/electronics12020302
by Yannick Uhlmann 1,*, Michael Brunner 2, Lennart Bramlage 2, Jürgen Scheible 1 and Cristóbal Curio 2
Reviewer 1: Anonymous
Reviewer 3:
Electronics 2023, 12(2), 302; https://doi.org/10.3390/electronics12020302
Submission received: 29 November 2022 / Revised: 23 December 2022 / Accepted: 25 December 2022 / Published: 6 January 2023

Round 1

Reviewer 1 Report

Electronics-2096732-peer-review-v1

Procedural and Reinforcement Learning based Automation Methods for Analog Integrated Circuit Sizing in the Electrical Design Space

By Yannick Uhlmann,* , Michael Brunner, Lennart Bramlage, Jürgen Scheible and Cristóbal Curio

 

This paper proposed a machine learning-based design methodology for analog integrated circuit sizing. The Reinforcement Learning is utilized to achieve this fully autonomous sizing approach. The effectiveness of the method is verified based on the agents’ learning behavior and performance evaluated on circuits of varying complexity and different technologies. The paper is well prepared. I have the following comments for your consideration.

 

1. In the introduction, I suggest the authors to simply introduce the meaning of Integrated Circuit sizing automation. E.g., What is the conventional solution to IC sizing except automation? What is the advantage of automation.

 

2. Why choosing the reinforcement learning in your method? There are many newly proposed machine learning methods and why others don’t work in your case?

 

3. The performance of the machine learning method depends on the diversity of the training dataset. How did the authors obtain and select the training dataset?

 

4. The authors can briefly describe the findings in the conclusion.

 

 

Author Response

Thank you for the feedback, we have addressed your review:

  1. The conventional way of analog IC design, which is either using textbook equations or the gm/Id method, is described in Section 1.1. This section also describes the gap between automation of digital vs. analog design. Where analog design is severely lacking behind, bottlenecking the flow of mixed signal chips. Thus, the advantage of automation is shorter time to silicon.
  2. Section 1.4 details the benefits of RL and why we believe it is an appropriate approach for analog IC sizing automation. An additional paragraph in Section 7.2 was added to emphasize again why this is our belief.
  3. In case of the procedural approach, Section 3 is dedicated to the dataset. Here we describe in detail how to generate the appropriate data, which features to transform and how to sample the data for efficient training. Unfortunately we cannot share the dataset, because the transistor data is protected by NDAs. For RL, there is no dataset. The data is collected during training and stored in a replay buffer, which is discarded after training is complete. The dimensions and parameters of action and observation spaces are described in Section 5.
  4. Due to multiple reviews addressing this issue, we have expanded the Conclusion substantially.

 

Reviewer 2 Report

I think the paper is well written and can be published in the journal taking into account the following sugestions:

1. What is the motivation in dividing the introduction into subsection. I think that is more elegant not including subsections, I reccomend removing the subsections in introduction.

2. I recommend to improve the motivation paragraph in introduction, better relating previous works with the purpose of the paper.

3. Please improve paragraphe between lines 194 to 203 for better show the alternative approach that you propose.

4. Please apropriately define the terms that appear in Figures 1 and 2, also in Equation 1.

5. please review the grammar of the entire document. In some parts you separate the subject from the verb with a comma, for example lines 229, 335

6. Please improve the description of results of Figure 5

7. Conclusion must based on findings of result section.  Please expand on the conclusions by showing this aspect better

8. Please review the journal template, I see that the authors do not follow it to mention Figures, Equations, Tables, author contributions, etc.

Author Response

Thank you for the feedback, we have addressed your review:

  1. Splitting the Introduction into subsections gives structure to the different topics and concepts we are trying to introduce for the two different approaches. Additionally, we would like to refer the reader to specific points in the introduction during the main text, this is only possible by using subsections.
  2. The motivation paragraph in the introduction is merely an opening sentence, giving context to the following subsections. Our motivation is spread across the following subsection (1.1 - 1.4), since it is related to different subjects. This has been specified now by clarifying it at the beginning.
  3. To avoid redundancy in the manuscript we refer the reader to Section 4 by the end of the paragraph, where the approach is explained in great detail.
  4. The missing definitions for variables and terms have been added where appropriate.
  5. Additional grammar reviews regarding punctuation were conducted.
  6. The description for Figure 5 has been expanded to further clarify what is shown.
  7. Due to multiple reviews addressing this issue, we have expanded the Conclusion substantially.
  8. Fig. and Tab. references were adjusted to conform with the template. The template does not explicitly specify how to reference Equations. The 'Author Contribution' block was generated by the submission form and included in the document as is. If this violates the template, what should be used instead?

 

Reviewer 3 Report

 

In this paper procedural and reinforcement learning based automation methods for analog integrated circuit sizing in the electrical design space is investigated. But article has serious flaws and research not conducted correctly. Many parts of paper previously published by authors and the plagiarism check index is higher than standard value. The paper can’t be accepted with this situation.

There are some comments, which are listed as follows:

 

-There are several typos for example:

In Abstract: Analog Integrated Circuit sizing is notoriously… There are several words with capital letters, which should be modified.

In table 3 and 4 the heading are written below the table and in the others tables heading are written above tables.

 

- Many parts of paper previously published in [R1], and copied without any changes, which restrict the novelty of manuscript and should be modified.

-The novelty of paper should be clearly emphasized.

-The Plagiarism check index is higher than standard value and should be modified.

- Equations which are not belongs to authors should be cited.

-Figures which are not belongs to authors should be cited.

-The abstract and conclusion should be modified with numbers and parameters improvement. Only quality report is not sound scientific report.

- The quality of figure 3 and figure A1 (a,b,c) should be improved. The written text should be readable in printed version or 100% zoom.

- In order to have better clearness a comparison table should be added and obtained results should be compared with some related works.

 

 

 

[R1] Yannick Uhlmann, Michael Essich, Lennart Bramlage, Jürgen Scheible, Cristóbal Curio. "Deep Reinforcement Learning for Analog Circuit Sizing with an Electrical Design Space and Sparse Rewards", Proceedings of the 2022 ACM/IEEE Workshop on Machine Learning for CAD, 2022

 

Author Response

Thank you for the feedback, we have addressed your review:

  • The capitalizations were due to the definitions of acronyms in this way, which were used here for consistency. This has been rectified by adjusting these definitions with the correct lower case spellings. The erroneous table captions have been fixed.
  • Section 5 has been revised to address this issue.
  • Clarification regarding the novelty of the work has been added to the conclusion in Section 7.2.
  • All equations are cited appropriately where they are referenced in the text.
  • All figures belong to the authors. Citations to prior work were added.
  • Due to multiple reviews addressing this issue, we have expanded the conclusion substantially.
  • Schematics have been scaled appropriately to increase readability as requested.
  • A direct comparison to related work might be inappropriate considering the difference in performance metric and action space. Never the less, a comparison, albeit under reservations, was added to the conclusion. Here, we compare the sample efficiency and navigational capabilities of the agents. To explain the numbers shown there and give them more context the comparison is not in a table but continuous text.

Round 2

Reviewer 3 Report

 

In this paper procedural and reinforcement learning based automation methods for analog integrated circuit sizing in the electrical design space is investigated. The authors have addressed most of my concerns and manuscript improved compared to the previous version, but there are still some modifications, which should be considered.

 

 

- The similarity is still high and paragraph which exactly copied from [R1] and other works should be modified.

-Quality of figures should be improved. Specially Fig.4,  The written text in the figures should be readable in the printed version or 100% zoom.

 

[R1] Khaleel SA, Hamad EK, Parchin NO, Saleh MB. MTM-Inspired Graphene-Based THz MIMO Antenna Configurations Using Characteristic Mode Analysis for 6G/IoT Applications. Electronics. 2022 Jul 9;11(14):2152.

 

 

Author Response

  • Considering the previous review, we assume [R2] is meant. We have revised Sections 5 and 6 to further address this issue. The only remaining similarities should be contained to the Introduction now. Since this section is a recap of related work and state of the art and not directly part our research, similarities should not be as critical there. It was our understanding, that for this journal extended versions of papers are acceptable, as long as there are at least 50% unpublished additions. If there are still any remaining issues regarding similarities to previous work, we require specific references to the offending paragraphs and sections in the manuscript, so we can address them adequately.
  • The scaling of Figure 4 has been adjusted, making it readable at normal print size.

[R2] Yannick Uhlmann, Michael Essich, Lennart Bramlage, Jürgen Scheible, and Cristóbal Curio. 2022. Deep Reinforcement Learning for Analog Circuit Sizing with an Electrical Design Space and Sparse Rewards. In Proceedings of the 2022 ACM/IEEE Workshop on Machine Learning for CAD (MLCAD '22). Association for Computing Machinery, New York, NY, USA, 21–26. https://doi.org/10.1145/3551901.3556474

Back to TopTop