Optimize Planning Heuristics to Rank, not to Estimate Cost-to-Goal

In imitation learning for planning, parameters of heuristic functions are optimized against a set of solved problem instances. This work revisits the necessary and sufficient conditions of strictly optimally efficient heuristics for forward search algorithms, mainly A* and greedy best-first search, which expand only states on the returned optimal path. It then proposes a family of loss functions based on ranking tailored for a given variant of the forward search algorithm. Furthermore, from a learning theory point of view, it discusses why optimizing cost-to-goal \hstar\ is unnecessarily difficult. The experimental comparison on a diverse set of problems unequivocally supports the derived theory.

Paper

References (61)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer ahr36/10 · confidence 4/52023-06-26

Summary

The paper highlights the limitations of the conventional method for learning heuristics in search algorithms, which involves imitating the optimal cost-to-go function. The authors present an alternative approach that focuses on optimizing the correct ranking of state pairs by the heuristic. Because of that characteristic, the authors argue that the resulting heuristic offers both theoretical and practical advantages compared to the standard approach.

Strengths

The paper presents an original idea that indeed can offer a significan advantage for planning methods. The presentation is mostly clear, with only minor problems.

Weaknesses

In my opinion, the claims formulated in the paper should be supported better. I'm not convinced that the reported results stem from the theory. Clearer discussion of theoretical advantages and disadvantages of various approaches, more detailed experiments including analysis of the learned heuristic, and comparison with some other baselines could be a significant improvement to the paper. Details of my concerns are discussed in the comments.

Questions

In general, I like the idea and the paper, although at this point it may need some improvements. At this point I'm not fully convinced that the improvement you show is due to the theoretical advantages (which should be clarified) or just a fancy heuristic that happens to work better than a certain imitation objective. Below I highlight a few specific comments. - l.111,117,157 Arguably a central expression, appears in 3 places, and every time is written differently. Correct it (or explain that it's actually not the same). - l.113-125 This explanation of the notation above is a lifesaver. - l.131 Does there exist at least one perfect ranging for every problem? Semms that h(s):=[-infty on a fixed optimal path, infty otherwise] works. What happens in quite a common case in which s_0 is not specified and the heuristic should work for any starting state? Does there always exist a perfect ranking in such case? - l.105-150 While I don't deny the soundness of the theory, it reads a little like explaining quite an intuitive idea ("the heuristic is strictly optimal iff in every step of the search the next node on a fixed optimal path receives the lowest value") with very complicated expressions. I'd consider changing the order and starting from intuition, moving to the formalities afterwards. The ultimate conclusion should be the expression in l.111, which I believe is the central object in this section. - l.161 Quite a strong assumption. - l.167 The relation between log(1+exp(x)) and [[x>0]] indeed seems loose. On convergence it will have the desired property indeed, but it doesn't exactly optimize the number of violations, rather a sum of violation volumes. How does it relate to your theory? An experimental analysis of the learned function would do. - l.180 I'd consider rewording the section title. Although after reading the section it makes sense, at first it seems like a general discussion of optimizing neural networks. - l.181-185 I like the idea, but it's not true that it is easier to learn a function from a bigger class. The core element is the learning signal. In case of cost-to-go, where we have actually a direct supervision, it is of much better quality than in case of a set of inequalities. Consider an example in which we have n states and we want to learn any monotonic function. Although there are much more functions satisfying f(s_{i+1})>f(s_i), learning fixed supervised targets should be faster. - l.186-196 "states off the solution path" -- you mean in the training dataset or in general? As far as I understand, your training dataset consists of solving paths only, so the issue applies to your ranking loss as well, doesn't it? When provided with off-the-path states, standard cost-to-go functions can also utilize them for learning (e.g. bootstrapping for Q-learning). The issue you describe is devoted to the particular training loss you consider, not learning cost-to-go in general. - l.203-207 I don't understand. Optimizing the proposed losses never considers dead-ends as they don't lie on the optimal paths. Do you consider any hand-modified heuristics? If you are worried about too high gradients, use Huber loss instead of L2. Why should lare values appear in the training? (5) is strictly supervised. - l.269 That's a shame, it would support your claims well as it makes a middle-ground between (3) and (5). - l.285 How costly is it to get this dataset? Is your method more efficient than SymBA* afterwards? Gathering optimal paths for training may be very costly, especially in the hardest domains. Thus, can you comment on the applicability of your claims, since you require a costly operation of gathering optimal paths for your training> Can your method, like standard imitation, learn also from sub-optimal data, which is often much easier to get? - l.286 Why a single problem instance? - l.322-342 If the test sets are harder than training sets, how does it relate to the theory in which you assume having optimal paths? - l.343-363 I'm not familiar with the complexity of some instances of the environments. Thus, it would be helpful to see some trivial baselines in the comparison, such as BFS or Dijkstra. Also, to showcase an advantage of your theory-driven approach, I'd consider comparing with a simple objective that learns to output 0 on the solving paths and 1 on other states outside the path. It fits the definition of optimality and doesn't seem to be stable or generalizable (or maybe?). Advantage over such a baseline would highlight the benefit of learning the ranking over heuristics. It's missing now. Yes, I know that you compare already agains many relevant baselines. - l.343-363 Also, the claims should be supported by some ablations. Is the learned heuristic correct? Does it learn to properly rank states, even in the training domains? What does it converge to, do the values resemble cost-to-go or some other interpretable quantities? What are the mistakes? Does the learned heuristic fit the theory of what it should learn? Are the paths obtained this way optimal? How do they relate to optimal paths in terms of length? How often only states on an optimal path are expanded? Can you learn a useful heuristic for a fixed domain with many possible starting points? - l.375 I'm not sure whether optimizing set of inequalities is easier than a direct supervision. And a few minors: - l.31,212,215,271 etc. When you refer to something, either in your paper or in supplementary, refer to the exact section. - l.40 Do you mean "state-of-the-art"? - l.48 I wouldn't say "always", you can demonstrate that it's better "in those eight environments". - l.50-72 I can't find a definition of g(s) or h(s) before you use it to explain your algorithm. Maybe in preliminaries? g(s) is defined below the algorithm, but h(s) is not. - l.116,135, and others: I feel confused whether "heuristic" refers to h or f. Please disambiguate the notation. - l.305 24000 gradient steps in 100 GPU hours? Seems quite slow, I'd suggest checking it.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

The limitations are not clearly stated, there should be a dedicated section for them. Societal impact is not an issue here.

Reviewer 458p6/10 · confidence 4/52023-07-02

Summary

Planning with (possibly learned) heuristic function is important for both planning and learning communities. The paper focuses on the forward search without expanding redundant search states in their forward process. With the development theorem, the authors demonstrated how the proposed concept work in various domains with A* and GBFS algorithms.

Strengths

- The concept of the focusing problem, wheares they seem to be simple, is interesting. The proposed theoretical discussion (e.g., sufficiency and necessity in Theorem 1) is clearly explained; the whole paper is well written and structured. - The proposed approach with $L_\star$ based on the two fundamental search techniques A* and GBFS work efficiently than counterparts.

Weaknesses

- Although the proposed learning-based approach is promising, it requires much learning time on PDDL domains (e.g., 17724 CPU hours). - The scales and variations of compared benchmark domains seem limited (e.g., traditional search problems and mazes).

Questions

The paper is very clearly organized, and I can focus on the content. The below items are come comments. - Learning-based heuristic functions could be promising for their generalization ability. However, sometimes they are argued that the required learning times (or resources) are too large. In the reported times with PDDL domains, it requires 17724 CPU hours. These results may cause some questions about the efficiency of the learning-based heuristic itself. - Do you discuss why the learning-based heuristic is promising, even though the training resource is required? - Can we compare planners without learning and planners with learning (i.e., the proposed method) in terms of total required resources? - Can the two solvers (A* and GBFS) have similar characteristics in terms of learning-based methods? The two solvers can be formulated similarly, as noted in L.93. Can we regard these methods as just variants of search using the function of the form in L.92, or can we have some important findings of these methods through the experiments? - Mathematical notation issues: - In Def. 1, some mathematical errors seem to occur in the second condition: see L.111 and L.117 to check the condition of $\wedge$. - In Eq.(6), the log() is not closed.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

I have no explicit concerns about the limitations.

Reviewer 7xum5/10 · confidence 3/52023-07-04

Summary

This paper investigate the computational advantage if a forward-search algorithm utilises ranking rather than optimising cost-to-goal. I think the contribution is good in general.

Strengths

1. The problem is formulated well. 2. The introduction of the contents is suitable. 3. The experiment section is good.

Weaknesses

1. The definition of "optimal ranking" is so important to this paper that it needs to be presented before first used. The authors should not use "(definition is below)" for this term. 2. I suggest the authors simplify some expressions. For example, at some points where strict rigorousness is not required, "minimise the number of expanded states"--> "efficient". 3. I cannot understand line 7 in the supplementary file. Is it grammatically correct? 4.Line 10 in the supplementary file, it is not "lastly". By the way, what is the connection between the Section 2 in the supplementary and the main text? I don't think there is a need to present this section. As long as "estimating the cost-to-go converges" is slower than the proposed scheme, the contribution of the main text is valid. Am I correct? 5.Can the author find a citation for the claim of Example 1? so that there is no need to present it in detail. I think this example broke the reading flow. 6. Why is the related work section at Sec 5? Why can't it be Sec 2? 7. I don't think Sec 4.3 presented rigorous statements. All things are subjective. Did the author write "conclusive remarks" here? I suggest the authors move them to the conclusion section. Don't be afraid that the algorithm section is too short. Minor points: 1. I didn't find the definition for the \wedge symbol in Def.1.

Questions

Please explain my comments listed in the weakness block.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

I don't think the paper has any negative societal impact. But I do suggest the authors polish their Sec 4.3.

Reviewer jMYD7/10 · confidence 4/52023-07-07

Summary

The paper considers the problem of using imitation learning for obtaining planning heuristics suitable for using variations of best-first search with a weighted sum of h() and g(), covering A* and GBFS. The paper gives both theoretical and empirical arguments for predicting a ranking of actions instead of predicting h*, the optimal distance to the goal. One argument in favour of ranking is that regression is a harder problem than regression. Another one is the observation that even with an optimal heuristic, the search can still expand more nodes than necessary. For that, in the case of symmetries, the proposed algorithm would choose a particular optimal solution among all the possible ones. The empirical results show the proposed approach to perform the best in A* in comparison with the others. A heuristic trained for GBFS performs well.

Strengths

- Interesting discussion of ranking vs value estimation. - Hard-to-achieve goal of expanding strictly nodes along the path. - Discussion on the pros and cons of the approach. - Including an honest commentary on sample complexity. - Promising empirical results.

Weaknesses

- The proposed approach does not account for flexible goals or multiple goals. - Limited discussion on the cases of symmetries and multiple optimal paths. - This is discussed but it could be analyzed further in the experiments.

Questions

- What happens in the case of symmetry or multiple paths? - For instance, suppose there are k discrete integer variables, that can be increased and decreased independently. Suppose the goal is to set them to 0. Optimal plans just interleave increasing the variables independently. What would the theoretical analysis and the empirical results look like? - This example might be unrealistic but it's related to problems with very flexible solutions like the PDDL benchmark Logistics. - How sensitive is the method to the distribution of problems and trajectories? - Is there any reason for it behaving differently across the proposed methods and the others?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The line of work is interesting, and there are certain domains where this method is relevant. The paper might benefit from exploring better those cases, given that the paper admits that there are cases where estimating the value might be better. It'd be good to discuss the stability of training for the different algorithms. The space used for introducing the search algorithm could be used for further discussion.

Reviewer ahr32023-08-17

Thank you for the response

- l.111,117,157 In l.117 there seems to be an extra conjunction, in l. 157 $s_j$ is changed to $s$ and the superscript of $\mathcal S$ seem to be missing $\pi$. While these are all details, they make the expressions less readable. - l.131 I mean that Theorem 1 provides an equivalent condition for a perfect ranking, but it is also important to know whether the definition is robust enough that a perfect ranking exists at all. As you said, we know it if the initial state is specified. I'm curious about a case when the initial state is not specified. Is it true that for every environment there exists a function that is a perfect ranking for every possible initial state? (it would be even better to understand how does it look like) Can you give an intuition why it holds, or a counterexample? Proving such a theorem would make your paper stronger. The trivial case of "1 on an optimal path, 0 elsewhere" is clearly too strong for the general case. Is your definition robust enough to cover all possible scenarios? P.S. I think no: if we have four states A,B,C,D, the goal state is A, and d(A,B)=d(A,C)=d(B,D)=1 and d(A,D)=d(B,C)=9 there is no ranking that is perfect for both C and D since they both expand to {A,B} but require different ordering. However, if you include the information about the last action executed, such a ranking always exists (the graph of optimal actions is well-defined, so you can estimate 0 there and 1 for other actions). While such an analysis may be out of the scope of this paper, I suggest thinking about it. - l.186-196 Can you also clarify whether your method suffers from this issue? I understand that you somehow use states off the solving path during training. Where do they actually come from? Could you refer to the place in the paper where it is explained, I can't find it now. - l.343-363 I acknowledge the results you provided. I think that such a discussion strengthens the presented results, I suggest including it in the paper, even in an appendix. I see that even in such a small example, the magnitude of the learned estimates is considerable. Do you somehow control it to ensure no blow-up?

Authorsrebuttal2023-08-17

Is this a solution to your example?

* Regarding the l.117. and l.157, we think we understand your point now and will easily fix it. We have not thought the differences between $s_j$ and $\mathcal{S}$ to be disturbing, but we are happy to know about them now and change it according to your suggestions to improve the readability. * Now, we understand your point and we think the function should exist if we do not take into the account its computational complexity (which is a helper to use the solver to identify optimal solution path). But we need to think about the formal proof and it would be probably a different work. Especially we need to think what can go wrong. We have tried to solve your example on the paper and we think the solution exist. we will talk exclusively about A*, because GBFS does not have a guarantee on optimal and therefore the quality of the solution does not matter. Imagine the heuristic values of states to be h(D) = h(C) = h(A) = 0, and h(B) = 4. Let's now simulate the behavior of A* for different starting states {B,C,D} **Starting at D:** A* will put to open list states A and B. The merit function $f$ will be: $f(A) = 9 + h(A) = 9$ and $f(B) = 1 + h(B) = 5.$ A* takes from open list state B, check if it is a goal (it is not) and expand it to A. When inserting A to open list, A* founds a better bath to A and updates the path to A to ((D,B), (B,A)). The open list contains only one state A, which is the solution state and optimal solution was found. **Starting at C:** A* will put to open list states A and B. The merit function $f$ will be: $f(A) = 1 + h(A) = 1$ and $f(B) = 9 + h(B) = 14.$ A* takes from open list state A, check if it is a goal (it is), and return solution path (C,A). **Starting at B:** A* will put to open list states A and C. The merit function $f$ will be: $f(A) = 1 + h(A) = 1$ and $f(C) = 9 + h(C) = 10.$ A* takes from open list state A, check if it is a goal (it is), and return solution path (B,A). Unless we understand incorrectly your example, the optimal ranking $h$ exists. * l.186-196 We collect states off the optimal path by expanding states on the optimal path (of a given problem instance). In our implementation, we do this when we are preparing mini-batches, as the complexity is trivial. We might not have discussed this in sufficient detail, since we considered this to be an implementation issue. We have believed the need to collect these states to be clear from the theoretical exposition of the loss. We now see that adding a sentence about this to the experimental settings might make it more clear. * l.343-363 The large values of heuristic values in the example were caused by the logistic loss convex surrogate, which is nonzero and strongly convex. A hinge loss would create a different value which might be way smaller. As we have mentioned on the L.182, the solution set is invariant to strictly monotone transformations (in case of A* algorithm, the set of transformation is smaller), therefore the optimization algorithm has picked a solution and did not care about size of heuristic values.

Reviewer ahr32023-08-19

> We will talk exclusively about A*, because GBFS does not have a guarantee on optimal and therefore the quality of the solution does not matter. As far as I understand, your theory covers all combinations of $\alpha,\beta$, which include GBFS and nearly-GBFS planners. In fact, when it comes to theory, such instances are more interesting since, as you observed in Example 1, for A* the existence of perfect ranking is easy -- it is the cost-to-go function (up to the tie-breaking issue). Thus, also in my example, the cost-to-go is a perfect ranking for A*. If $\alpha,\beta > 0$, we can always rescale $h(s)$ to pretend that $\alpha=\beta$. However, this is not the case with GBFS ($\alpha=0$). In my example, I think there is no optimal ranking for GBFS, because of the reasons I provided.

Authorsrebuttal2023-08-19

We think that for GBFS the optimal ranking exist as well. For the case you have provided it will be as follows: h(d) = 2, h(b) = h(c) = 1, h(a) = 0, which it is equal to the number of nodes that needs to be expanded to reach the goal. Similarly as above, we can verify that GBFS is optimally efficient as it will find the solution while expanding the minimal number of nodes. **Starting at D**: GBFS will put to open list states A and B with heuristic values h(A) = 0 and h(B) = 1. GBFS takes A from the open list and checks if it is a goal state. Since A is goal state, GBFS terminates returning path (D,A) as a solution. **Starting at C**: GBFS will put to open list states A and B with heuristic values h(A) = 0 and h(B) = 1. GBFS takes A from the open list and checks if it is a goal state. Since A is goal state, GBFS terminates returning path (C,A) as a solution. **Starting at B**: GBFS will put to open list states A and C with heuristic values h(A) = 0 and h(C) = 1. GBFS takes A from the open list and checks if it is a goal state. Since A is goal state, GBFS terminates returning path (B,A) as a solution. The optimal ranking for all states for GBFS therefore exist, as it always finds the solution while expanding minimal number of states. But the ranking function does not guarantee that the GBFS will find optimal solution. If we are not mistaken, we have not claimed in the paper that GBFS with optimal ranking function is guaranteed to find optimal solution. We claimed it will find the solution while expanding minimal number of states. We also said (with proof in suplementary) that if all actions have same cost, then optimal ranking for A* is optimal ranking for GBFS, but this is not the case of this problem. This example nicely shows that optimal ranking is **search dependent**. We will try to explain this difference more explicitly and add this example into the supplementary, if you agree.

Authorsrebuttal2023-08-19

I got it

We now understand your point and your are right. For GBFS, there is no optimal ranking for all states. This is because the definition of *optimal ranking* requires the inequalities to hold for the optimal path. Thus you are right that the heuristic does have to exist for all initial states and your example is correct. The problem can be fixed if the definition of optimal ranking is changed to consider two cases: $\alpha > 0$ and $\beta > 0$ and $\alpha = 0$ and $\beta > 0$. In the first case, the optimal ranking should be defined with respect to the "optimal path" (this is the A*), in the latter it should be path with minimal number of expansions" for GBFS, because with GBFS the user wants to found a solution, but should care about its optimality. With this change, the heuristic function will exist for the above example. Thank you very much for this discussion. It was very enlightening and we appreciate your input.

Reviewer ahr32023-08-20

Thank you for the discussion and clarifications. I acknowledge that by increasing my score.

Authorsrebuttal2023-08-21

Thank you very much. We appreciate the discussion.

Reviewer 458p2023-08-17

Thank you for clarifying the concerns, particularly for total costs.

I appreciated the updates from the authors. I may misunderstand something about the learning points. However, the responses above clarified some of them, and I understand some argued difficulty in considering the total development costs (for classical planners). For the findings (of A*-based and GBFS-based learning planners) in the paper, I agree that the findings are interesting for the literature. As I'm not so frustrated with the notations, but it seems to be better for improving the readability (as other reviewers commented). Any way, thank you for your comments and contributions.

Reviewer 7xum2023-08-17

Thank you for the rebuttal. Sorry for my delayed response. Please see my response below. 1. Rigorous issue: I don't think the authors understood me. The authors should know that when strict rigorousness cannot be established by non-clumsy wordings (here it is the case), then it is better to present the thing such that what the reader intuitively thinks is exactly what you want to present. As a counterexample, there are many "efficient" in the paper, but the authors should not expand all of them to "minimise ...". 2. Description in supplementary issue: I don't think the authors need to say things like line 7. (In my opinion supplementary did not do a good job actually.) 3. I agree with the authors' opinion that "ranking" is probably easier to estimate than "cost-to-goal". I did search related works for a while (therefore I'm late for the response, apologise again for the delay) and found one regarding this: [1] Yang T, Huang L, Wang Y, et al. Efficient search of the k shortest non-homotopic paths by eliminating non-k-optimal topologies[J]. arXiv preprint arXiv:2207.13604, 2022. Given the fact that the citations are indeed old, I hope the authors can look for more recent papers. 4. Example 1: Yes, I suggest remove it. 5. Position of related works section: After carefully re-read this section, I think does not explain the art of the related topics well. I'm aware that it is difficult to expand (as I mentioned above). So my suggestion is that the authors first revise it, and we discuss its position later. 6. The authors have understood my opinion. Sec 4.3 has been at the place for discussion. Therefore, if a proven statement is presented, then please put it into the conclusion section (to recall the contribution). If a conjecture is raised, then put it to a subsection entitled "discussion". The authors have clearly distinguished the role of each paragraph in the rebuttal.

Authorsrebuttal2023-08-18

We really appreciate your suggestions how to improve the writing and organization of the manuscript. There are few things we do not fully understand in your response. 1. Can you be more specific in which aspect the supplementary did a poor job? We add the point about goal-awareness, since this is a difference between estimating cost-to-goal and ranking. We wanted to be honest that ranking losses as they does not make the heuristic goal-aware. 2. From your response, we did not figured out, how the reference [1] is related to the fact that ranking is easier than regression. This is cause that we not at all familiar with the non-homotopic path and it is going to take us while to understand it. We know that our citations are old, nevertheless they very well carry the message. We wanted to find something recent, but were so far not successful. With that respect, we are grateful for your reference. 3. The related section is indeed brief, because the conference paper has a tight page limit. And although it would be really nice to write a longer text about the prior art, it would not fit the page limit, and honestly, it would be a different paper (something more like a survey). We therefore cited all works we were aware of that proposed something related to out work. We would be happy to know more details of how would you improve or what you do not like on the related work section. If we understand correctly, your criticism is related to the writing and structural aspect of the paper. The suggestions makes a lot of sense and they will help the paper, but they are relatively easy to incorporate and we would be more than happy to implement them for the final version of the paper. We would also like to ask, if there is something regarding the technical and experimental contribution of the work that you do not like? Thanks a lot in advance for your response and time. We appreciate your comments, as they are generally good advices in scientific writing.

Reviewer 7xum2023-08-19

Thank you for the response. 1. Related works issue: In my opinion, the proposed algorithm is trying to optimally rank the nodes in the priority queue of a searching-based algorithm. The precise estimation of heuristics of the nodes in the queue is a sufficient condition of obtaining the optimal ranking of the nodes in the queue (which is also a sufficient condition of the optimal choice of the next best node to be expanded). Hence, instead of estimating heuristics, estimating the ranking is an easier task. This is true. On the other hand, in [1], the algorithm compared the relative optimality of the searching branches in non-homotopic topological routes, which is very similar to the definition of "ranking". The elements that were compared in [1] are not X-Y locations of the robot but different homotopy classes of paths. I also think the authors might not be familiar with classic planning which is OK. But given the lack of related works I hope the authors re-do a careful searching of relative literatures. 2. Writing: I don't think the writing issues in this paper are easy to fix. If they are, then the authors should have fixed them before the submission. Given the fact that the authors did not even show me anything about how the manuscript is improved during this rebuttal phase, I have no further comments regarding this.

Authorsrebuttal2023-08-19

Thanks for clarification of 1, we will give it a try. How would you like us to show that we are happy to incorporate your suggestions to the manuscript, when we cannot update the manuscript during the rebuttal phase?

Reviewer jMYD2023-08-20

Thank you for your answers. The example is very insightful. After reading this and other responses, I'm increasing my rating to accept. Apologies for the bad writing of my question, 'Is there any reason for it behaving differently across the proposed methods and the others?'. I meant to add details to the question you answered 'How sensitive is the method to the distribution of problems and trajectories?', so it could have been phrased as "Is there any reason for the algorithm to behave differently across different datasets?". I'm satisfied with this answer: 'The trajectories that are more frequent in the training set will be preferred. '

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC