Applying Decision Transformers to Enhance Neural Local Search on the Job Shop Scheduling Problem

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper introduces a novel approach using Decision Transformer (DT) and suggests its potential to learn more effective exploration strategies compared to traditional reinforcement learning models. Additionally, it experimentally validates the feasibility of applying Transformer-based models to JSSP optimization. I list some suggestions to enhance the quality of this manuscript.

 

1. The manuscript does not provide sufficient theoretical or experimental comparisons to demonstrate why Decision Transformer (DT) is superior to existing Deep Reinforcement Learning (DRL) models.

• For instance, comparative experiments with established RL methods such as Proximal Policy Optimization (PPO) or Deep Q-Network (DQN) are not included.
• If traditional RL models achieve performance similar to DT, the necessity of adopting DT may become less evident. Therefore, it is important to clearly validate its advantages over these methods.
• If such experiments were not conducted, the reasoning behind this omission should be explicitly stated.

 

2. The manuscript mentions that DT has a longer inference time compared to NLS, but a detailed quantitative analysis of computational cost is lacking.

• A more systematic evaluation, including GPU usage, memory consumption, and execution time, would help assess the practical feasibility of DT.
• This aspect is particularly important for real-world industrial applications, where high computational costs may pose practical challenges for real-time scheduling.

 

3. The dataset is generated using a Neural Local Search (NLS) teacher model and NLS-based environment modeling, raising concerns about potential dataset bias.

• The manuscript should provide a clear explanation of why this specific approach was chosen over other alternatives.
• If bias mitigation strategies were applied, they should also be explicitly described and justified.

 

4. The study employs an offline learning setup, where DT is trained on pre-collected trajectories from NLS. However, the manuscript does not provide a clear theoretical justification for why this approach is more suitable for Job Shop Scheduling Problems (JSSP) than an online RL approach.

• It is necessary to explain why offline learning is preferable for this problem setting and whether it offers significant advantages over traditional online RL methods in this domain.

 

5. The manuscript should provide a stronger explanation of why DT is particularly well-suited for JSSP. Rather than focusing solely on technical aspects, it would be beneficial to discuss the unique characteristics of JSSP and how DT aligns with these challenges.

• If DT provides specific advantages over conventional RL algorithms in the JSSP domain, these should be clearly articulated.

Author Response

Please see the attachment!

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

You have proposed strategies to solve the problem of job shop scheduling problem.

I have some questions as below.

1, in 1 Introduction,show the current problems and what problems you have solved.

2, you have developed a method for training the decision transformer (DT) algorithm instead of decision transformer method,so change your paper’s title.

3, it is better to replace 3. Related Work with 2. Preliminaries.

4,in 3. Related Work, you should cite more papers published in 2024.

5,in 5.1.Results on Taillard Benchmark, one dataset is not enough, add more.

6,in Table.3, analyze why DT perform worse in some case.

7,show you innovation more deeper.

Thank you.

Comments on the Quality of English Language

need to be improved.

Author Response

Please see the attachment!

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper has introduced a method to train the Decision Transformer (DT) algorithm on search trajectories from a trained Neural Local Search (NLS) agent, enhancing decision-making sequences for the Job Shop Scheduling Problem (JSSP). This proposed method results in superior local search strategies albeit requiring longer computation time.

The manuscript is well written and the provided experimental results show improved performance over the baseline NLS algorithm (Table 3 and 4). However, the comparison with state-of-the-art works is limited to NLS and only for Taillard benchmark dataset. The reference work for NLS (Falkner [18]), for instance, provides very detailed results on Uchoa benchmark as well. Moreover, it has been mentioned that another method “Variable Neighborhood Search (VNS)” [27] beats NLS under certain scenarios (smaller instances). Thus, it is advisable to expand the experimental section to include more comparisons with recent reference works.

Author Response

Please see the attachment!

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I would like to commend the authors for their thorough and thoughtful revisions. The manuscript has been significantly improved, successfully addressing the comments raised in the previous review. The authors have clarified key aspects of their study, enhancing the clarity and depth of the discussion. Notably, the revisions have effectively strengthened the explanation of core research contributions and their relevance to the broader field.

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for your modifications.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have adequately answered the questions raised previously.