Novel Bio-Inspired Physics-Based Learning and Evolutionary Guidance for Dynamic Multi-Objective Cold Chain Routings

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper proposes H-MODRL, a hybrid framework combining genetic algorithms, chaotic sparrow search, and physics-informed multi-objective proximal policy optimization for agricultural cold chain logistics routing. The manuscript has significant structural, methodological, and presentation issues that must be addressed before it can be considered for publication.

1- The paper contains a duplicate section heading: Section 2  and Section 3 

2- Please check and correct the start of section 4 "Datasets. Unlike ..."

3- Figures 10, 11, 12, and 13 all carry captions referencing "Scenario 2" even though they appear in Section 4.4, which covers mountainous terrain (Scene 3). Please check and correct if applicable.

4- The authors state they construct a custom heterogeneous terrain dataset from real Chinese geospatial data, yet provide no information about data sources, preprocessing steps, or public availability. Please give more information.

5- The paper promotes O(1) dynamic disruption lookup and millisecond-level re-planning as key engineering contributions, yet provides no runtime measurements or latency benchmarks comparing H-MODRL against baselines. Claims of real-time capability require more empirical validation  especially for larger-scale scenarios.

6- No sensitivity analysis on reward weights w₁–w₄. Given that weight selection critically determines the Pareto trade-off behavior in scalarized multi-objective RL, a dedicated weight sensitivity experiment is necessary.

7- Several sentences are excessively long and should be broken up for readability.

8- Remove a) b) c) points from the abstract and replace them with more smooth writing.

9 - The Smart Repair Operator in Stage 1 is described qualitatively but no pseudocode or formal algorithmic description is provided. 

10- How you select (Equation 2) the decay parameter λ_p = 0.002 ?

11- Figures 14–16 are all captioned identically as "Scenario 1 – Hyperparameter optimization" despite showing Scenarios 1, 2, and 3. This must be corrected.

Comments on the Quality of English Language

The English writing can be improved

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for submitting your manuscript on a hybrid multi-objective deep reinforcement learning framework for agricultural cold chain logistics. The topic is important, and your attempt to integrate bio-inspired optimization with physics-informed learning is ambitious. However, after careful evaluation, I am unable to recommend this manuscript for publication in its current form.

Below are the main issues that need to be addressed.

1. Clarity and Readability

The manuscript is very difficult to follow. Many sections contain long, dense sentences with multiple technical concepts introduced simultaneously, often without sufficient explanation.

• Key ideas (e.g., the role of each stage in H-MODRL) are not explained intuitively
• Mathematical formulations are presented without adequate interpretation
• Transitions between concepts are abrupt

Suggestion:

Revise the manuscript to prioritize clarity:

• Use shorter sentences and simpler structure
• Provide intuitive explanations before formal equations
• Clearly explain the purpose and effect of each component in the framework

2. Methodological Over-Complexity and Lack of Justification

The proposed framework integrates many components:

• GA + EDF initialization
• Chaotic SSA + LNS guidance
• Physics-informed PPO
• APSP + hash indexing

However, the manuscript does not clearly justify:

• Why each component is necessary
• How each component contributes to performance improvement

Critical missing element:

• Ablation study

Without this, it is not possible to determine whether the performance gains come from:

• the hybrid structure,
• the physics-informed reward,
• or only a subset of components.

Suggestion:

Include systematic ablation experiments removing or modifying:

• mapping
• LNS operator
• physics-informed reward
• guiding set

3. Structural and Organizational Issues

There are several structural problems:

• Section duplication (e.g., “Problem definition and physics-based modelling” appears twice)
• Inconsistent section flow between modeling and methodology
• Overloaded sections combining too many ideas

Suggestion:

• Reorganize the paper into a clearer structure:
• Problem formulation
• Methodology (with clearly separated stages)
• Experiments
• Ensure section titles and numbering are consistent

4. Experimental Validation

While the experimental section is extensive, it lacks important elements:

Missing:

• Statistical significance analysis (e.g., confidence intervals, hypothesis testing)
• Clear description of dataset construction and reproducibility
• Fairness discussion for baseline comparisons

Unclear:

• Whether all baselines are equally tuned
• Training budget comparability for DRL methods

Suggestion:

• Add statistical analysis
• Provide more details on datasets and implementation
• Ensure transparent and fair baseline comparisons

5. Overstated or Insufficiently Supported Claims

Some claims appear too strong or insufficiently justified, for example:

• “O(1) complexity”
• “millisecond-level real-time re-planning”

These claims require:

• clearer definition of scope
• empirical or theoretical support

Suggestion:

Moderate such claims or provide rigorous justification.

6. Language and Writing Quality

The manuscript contains multiple issues:

• grammatical errors
• incorrect word usage (e.g., “contents” instead of “contains”)
• inconsistent tense and phrasing

Suggestion:

A thorough language revision (preferably by a fluent English speaker or professional editor) is necessary.

Final Recommendation

The manuscript addresses an important problem and proposes an interesting hybrid framework. However, due to significant issues in clarity, methodological justification, experimental validation, and presentation, it is not suitable for publication in its current form.

I encourage the authors to substantially revise the manuscript, focusing on:

• simplifying and clarifying the presentation
• strengthening experimental evidence (especially via ablation studies)
• improving structure and writing quality

A thoroughly revised version could be reconsidered in the future.

Comments on the Quality of English Language

The manuscript contains multiple issues:

• grammatical errors
• incorrect word usage (e.g., “contents” instead of “contains”)
• inconsistent tense and phrasing

Suggestion:

A thorough language revision (preferably by a fluent English speaker or professional editor) is necessary.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript addresses an important problem in agricultural cold chain logistics: dynamic routing under time windows, freshness loss, carbon emissions, and uncertain disruptions. The proposed H-MODRL framework combines heuristic initialization, swarm-based guidance, and multi-objective PPO, and the experimental results show clear performance advantages over several baselines. The topic fits the scope of bio-inspired optimization and engineering AI, but several parts should be clarified or strengthened before publication.

1. The main contribution should be stated more clearly. The current framework includes GA, EDF, C-SSA, LNS, MO-PPO, APSP, hash indexing, freshness modeling, and carbon modeling. These components are useful, but the manuscript should better explain which part is the core innovation and how each module contributes to the final performance.
2. The role of reinforcement learning needs clearer explanation. Since the method uses heuristic initialization and an elite guiding set before PPO training, it is not clear how much of the improvement comes from MO-PPO itself. 
3. The term “physics-informed” should be used more carefully. The manuscript incorporates Arrhenius-based freshness decay and carbon-emission models into the reward and objective functions. This is reasonable, but it should be described as physics-based reward or mechanism-informed modeling unless the learning process itself is directly constrained by physical laws.
4. The claim of O(1) dynamic response should be refined. Hash indexing can support constant-time disruption lookup, but the complete re-planning process involves route decision-making and policy inference. The manuscript should distinguish between O(1) event query, fast state update, and total online decision time.
5. The experimental setting needs more detail. The paper states that the scenarios are based on real geographic information, but the data generation process, disruption simulation, customer demand setting, and time-window construction are not sufficiently transparent. 
6. The comparison with baselines should be made fairer and easier to interpret. Some baselines may not use the same precomputed distances, disruption indexing, or repair mechanisms. The manuscript should specify which engineering accelerations are shared across methods and which are exclusive to H-MODRL.
7. The ablation study should be strengthened. At minimum, the paper should report the effects of removing GA initialization, removing C-SSA/LNS guidance, removing the physics-based reward, and removing the elite guiding term. 
8. The freshness model parameters should be justified. The values of decay coefficient, temperature sensitivity, and temperature-control failure factor directly affect the optimization results. The manuscript should explain whether these values come from literature, empirical data, or simulation assumptions.
9. The figures are informative, but some captions and labels need correction. For example, several figures in the mountainous terrain section are still labeled as Scenario 2. These should be revised carefully.
10. The manuscript contains repeated section titles and some language issues. Section 3 is titled “Problem definition and physics-based modelling,” which repeats Section 2. Some sentences are long and should be simplified for readability. A careful language revision is needed.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you for handling in detail all of my review points and recommendations. The paper can be accepted in its current form.

Comments on the Quality of English Language

The English writing can be improved

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed all my concerns, and the manuscript is now ready for publication.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The revised manuscript has improved substantially. The authors have addressed most of the previous concerns by clarifying the three-stage contribution, replacing “physics-informed” with “physics-based”, refining the online response claim, adding more details on the experimental setting, and providing an ablation study. The revision is generally acceptable, but several points still need minor clarification before publication.

1. The contribution statement is clearer than before. However, the method is still mainly a hybrid integration of GA-EDF, C-SSA-LNS, and MO-PPO. The manuscript should avoid overstating the method as a fundamentally new DRL framework. The contribution is better described as a well-organized hybrid optimization framework for dynamic cold chain routing.
2. The new ablation study is useful and improves the credibility of the results. However, the physics-based reward and the elite guiding term are not fully isolated as separate ablation factors. The authors explain the difficulty of completely removing the physics-based part, but this limitation should be stated more clearly in the ablation section itself, not only in the limitation discussion.
3. The online response discussion is now more reasonable. Still, the manuscript should clearly distinguish optimization quality from online inference latency. Since some acceleration components are specific to H-MODRL, the comparison with baseline algorithms should not imply that all methods are evaluated under the same online deployment conditions.
4. The experimental setting is better described, but reproducibility remains limited. The authors mention OSM, traffic-weather data, Amap data, and industry reports. More concrete information should be provided where possible, such as the selected regions, parameter ranges, disruption-generation rules, and whether the dataset or code can be made available.
5. The biomimetic explanation has been expanded, but some parts are too descriptive. The analogies between algorithmic stages and biological systems should be shortened. The paper should keep the focus on the routing problem, optimization mechanism, and experimental evidence.
6. The manuscript should be checked again for formatting, figure labels, equation notation, and terminology consistency. The revision already corrects several previous problems, but a final careful proofreading is still needed.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf