Learning to Discover Skills through Guidance

In the field of unsupervised skill discovery (USD), a major challenge is limited exploration, primarily due to substantial penalties when skills deviate from their initial trajectories. To enhance exploration, recent methodologies employ auxiliary rewards to maximize the epistemic uncertainty or entropy of states. However, we have identified that the effectiveness of these rewards declines as the environmental complexity rises. Therefore, we present a novel USD algorithm, skill discovery with guidance (DISCO-DANCE), which (1) selects the guide skill that possesses the highest potential to reach unexplored states, (2) guides other skills to follow guide skill, then (3) the guided skills are dispersed to maximize their discriminability in unexplored states. Empirical evaluation demonstrates that DISCO-DANCE outperforms other USD baselines in challenging environments, including two navigation benchmarks and a continuous control benchmark. Qualitative visualizations and code of DISCO-DANCE are available at https://mynsng.github.io/discodance.

Paper

References (59)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer oWju6/10 · confidence 4/52023-06-14

Summary

This paper proposes an algorithm called DISCO-DANCE for unsupervised skill discovery in RL. The algorithm augments the mutual information reward of DIAYN with a gudiance reward. The guidance reward encourages indistinguishable or unconverged skills to follow skills that can potentially visit under-explored states, such that the over all skill collection can have broader state coverage. The authors conduct experiments on 2d navigation, Ant maze, and DMControl suites by evaluating the state coverage and the fine-tuned downstream task performance, showing that DISCO-DANCE outperforms baselines in most of these benchmarks.

Strengths

+ Illustration of the main idea (i.e., find and follow a guide policy) is clear (in particular fig.1) and well-motivated. + The authors provide open-source code and extensive experiment details such as hyperparameters and resource requirements, which make the results presented in this paper reproducible. + The appendix provides extensive discussions about the algorithm's limitation and comparison with other baselines, which can help reader understand the proposed algorithm deeper. + Experiment results are promising.

Weaknesses

+ The authors tend to address the problem unsupervised discovery in complex environments where existing methods are no longer effective, but experiments are mainly conducted on common benchmarks. I acknowledge that several maps in the navigation task are challenging, but the locomotion tasks (AntMaze and DMC) are not. Actually, baseline algorithms can outperform the proposed algorithm in terms of downstream task performance (see Appendix G).

Questions

+ The authors mentioned that new skills are added if existing skills are mostly converged. In my opinion, the skill-latent-conditioned policy is an MLP with states and latents concatented as the input. If a new skill is added, the input dimension is also changed. How do you add a new skill? Do you perform a neural network surgery or specify a sufficiently large number of skills at the beginning of training? + Correlated with the above question, since DISCO-DANCE adds new skills during training, are the total number of skills trained by DISCO-DANCE and baselines the same in experiments? + In the first sentence of bstract, I think diminished bonuses (e.g. DIAYN) are not equivalent to "penalties". + What is the x/y-axis of fig.3(b)? + Fig.6 is wierd. Why not present, e.g., the success rate over 100 trials averaged over 20 seeds with error bars? + In Algorithm 1, the guide skill z* is not defined if the "most skills are discriminable enough" condition is not satisfied. + Many meaningful results are presented in the Appendix (the base RL algorithm, additional ablation studies). I think they should be linked in the main body such that readers can have a sense.

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

3 good

Contribution

3 good

Limitations

+ The authors have addressed several main limitations in appendix I. + The selection of the guide skill depends on the final state visited existing skills. It does not seem to be a general solution to select guide skills even for state-based environments.

Authorsrebuttal2023-08-16

Kind reminder for discussion

Dear Reviewer oWju (reviewer 5), We deeply appreciate the time you've dedicated to reviewing our paper. While we believe we've addressed the concerns you raised, we have yet to receive your feedback. We feel that further discussion and feedback would greatly enhance the quality of our paper. We kindly request your continued engagement and insights. Best, Paper 12428 authors

Reviewer EWQN6/10 · confidence 3/52023-07-06

Summary

This paper proposes a new unsupervised skill discovery method based on the guidance of exploration. A policy is conditioned on a latent skill variable like in prior work. The method first starts by identifying a "guide" skill variable that is likely near unexplored states, the novelty of unexplored states is measured by the density of random walk arrival states started from the terminal states of each skill. It then trains other skills that explore the vicinity of terminal states region from guide skill.

Strengths

1. This paper analyzes the limitations of state coverage in previous skill discovery methods and provides nice motivation for the proposed method. 2. The empirical result in mazes and control tasks are promising.

Weaknesses

1. Measuring the density of the state distribution via generating random walk arrival states from terminal states is not sample efficiency. 2. The method needs to select a guide skill that is most adjacent to the unexplored states. In the bottleneck maze tasks in Figure 3, what if all skills including the guide skill cannot pass the first room? Will this method also encourages effective exploration?

Questions

1. To select guide skills for exploration, the proposed method assumes the terminal state is resettable and needs to measure the density of the state by generating about $O(PRM)$ random walks arrival states from terminal states as section 3.1 mentioned. I wonder whether is there any more sample-efficient and easy ways to do that. e.g. Is it equivalent to picking a guide skill by directly estimating the density of terminal states sampled from history episodes or fixed-length horizon? if the skill is discriminable (MI reward is well optimized), the terminal states from the same skill may locate in a subregion of (unexplored) states. 2. What does the y-axis mean in Figure 3(b?

Rating

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

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

N/A

Authorsrebuttal2023-08-16

Kind reminder for discussion

Dear Reviewer EWQN (reviewer 4), We deeply appreciate the time you've dedicated to reviewing our paper. While we believe we've addressed the concerns you raised, we have yet to receive your feedback. We feel that further discussion and feedback would greatly enhance the quality of our paper. We kindly request your continued engagement and insights. Best, Paper 12428 authors

Reviewer EWQN2023-08-18

Response

Thank the authors for the clarification and additional results in Figure b. There are no major concerns for me. I increase the score to 6.

Authorsrebuttal2023-08-18

We deeply appreciate your constructive feedback and updating the score. We will ensure that this clarification is adequately reflected in the revised manuscript.

Reviewer b5664/10 · confidence 4/52023-07-06

Summary

In this submission, the authors propose an unsupervised skill discovery method called DISCO-DANCE. It samples *guide skills* with random walk processes that start from the terminal states of a set of skills and use them to guide less discriminable or new skills toward those guide skills so that they can reach unexplored areas more easily. They test their method in two navigation environments (2d maze and Ant maze) and Deepmind Control Suite and compare the state space coverages and performances with the baselines.

Strengths

- In terms of orignality, the main idea of this work to find a reachable state that is close to the unexplored region and to expand the skill set based on it is novel to some degree. - The manuscript is mostly clear and easy to follow. Also, the concept figure effectively provides the intuition behind the method. - The state space coverage problem is an important aspect of unsupervised skill discovery.

Weaknesses

- The exploration issue with the mutual information (MI) objective could be more than what is described in this work. In theory, the MI objective is not supposed to contribute to the exploration meaningfully, especially in continuous control environments (Park et al. [21]), which can make this method mostly rely on the random walk processes for its exploration. - I believe one important weakness of this submission is the random walk process. The manuscript mentions that the rise of the environmental complexity makes existing skill discovery methods less effective and motivates this work, but ironically, in complex environments (e.g., with high-dimensional state spaces), random walk would be one of the main bottlenecks in encouraging exploration. In such environments, this algorithm could require a large number of iterations. - In terms of writing, I think it is not very fair to call the state spaces of the environments used for the benchmark *high-dimensional*. They are higher-dimensional compared to the 2D maze environment, but labeling them high-dimensional in general may not be a good standard for the field.

Questions

- Do you think there could be an alternative exploration strategy other than the random walk process?

Rating

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

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 authors state some limitations of the proposed method (difficulty in high-dimensional state spaces and stochastic environments), but I encourage the authors to consider taking the points I listed in the Weaknesses section into account.

Authorsrebuttal2023-08-16

Kind reminder for discussion

Dear Reviewer b566 (reviewer 3), We deeply appreciate the time you've dedicated to reviewing our paper. While we believe we've addressed the concerns you raised, we have yet to receive your feedback. We feel that further discussion and feedback would greatly enhance the quality of our paper. We kindly request your continued engagement and insights. Best, Paper 12428 authors

Area Chair q19X2023-08-18

Dear reviewer b566, you have raised issues on MI and on the random walk. May I ask you to check out the rebuttal on this (and the other reviewers) and to tell us (and the authors) what you think? Best, AC

Reviewer adLT7/10 · confidence 4/52023-07-06

Summary

This paper addresses the challenge of learning diverse skills in unsupervised reinforcement learning by introducing a method to selectively guide candidate skills to areas of the state space with low coverage from the current set of skills. Rather than relying directly on mutual information maximization like many skill-learning approaches do, this paper proposes an algorithm for selecting guidance skills and optimizing apprentice policies through a combination of mutual information rewards and a guidance reward. The guidance reward is high when the apprentice skill reaches similar areas of the state space as the guidance skill, which can help direct the apprentice skills to previously unexplored areas. The paper shows that the proposed method, DISCO-DANCE, can cover more of the state space than alternate approaches by evaluating in maze navigation environments. The paper also provides ablation studies to help understand what each part of the method contributes to the overall performance.

Strengths

The paper is clearly written, with each aspect of the DISCO-DANCE algorithm explained and compared with other methods from the literature. The contribution, both in terms of empirical results and novelty of the proposed method, is well defined. Diagrams were effectively used to make the approach intuitive. In addition to describing the method well, the paper evaluates the experimental performance thoroughly, comparing it to several similar methods on several domains. The experiments use both state coverage and fine-tuning performance, and show that DISCO-DANCE performs well.

Weaknesses

The main weakness of the paper mostly involves my uncertainty around the questions in the next section. I hope that the rebuttal can help clear up confusion.

Questions

1. Appendix F describes the additional costs of the random walk process, and proposes an efficient method. Are the additional steps used to perform random walks, even for the temporary buffer, included as environment steps in the overall training budget? If not, this seems an unfair advantage given exclusively to DISCO-DANCE when comparing to other methods. 2. The hyperparameters used for UPSIDE (from Appendix B) seem to limit the number of learned skills to 8 while DISCO-DANCE seems to start with more skills and has the possibility to expand. Can you comment on this choice? Wouldn't a higher number of skills for UPSIDE reduce the described fine-tuning difficulty by allowing more coverage? 3. DISCO-DANCE seems to require skills to reliably end in a terminal state. How are those states tracked and determined when the skills are still being trained?

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

Limitations are explicitly described and claims are not overly grand.

Authorsrebuttal2023-08-16

Kind reminder for discussion

Dear Reviewer adLT (reviewer. 2), We deeply appreciate the time you've dedicated to reviewing our paper. While we believe we've addressed the concerns you raised, we have yet to receive your feedback. We feel that further discussion and feedback would greatly enhance the quality of our paper. We kindly request your continued engagement and insights. Best, Paper 12428 authors

Reviewer gakY3/10 · confidence 5/52023-07-08

Summary

The paper introduce a reward shaping approach called DISCO-DANCE to enhance exploration in unexplored states in the context of latent variable skill learning in self-supervised RL context. The approach consist in defining a guide skill with the highest potential for reaching unexplored states for a given environment, then select unconverged skills and incentivize them to follow the guide skill expressed at terminal state, helping them bypass state regions with low rewards. Finally, the approach disperse the skills to maximize their distinctiveness, resulting in a set of skills that cover a wide range of states. DISCO-DANCE fills the pathway to unexplored regions with positive rewards. The approach is compared in navigation and locomotion scenarios surpassing some previous methods in terms of exploring the state space and performing well in navigation tasks in 2D mazes and Ant mazes.

Strengths

The paper is well written and explain rather well his approach. The problem of skill learning with self-supervised learning is important and actual. The justification is reasonably clear, maybe it would have been interesting to discuss more the difference between environment were most degrees of freedom are part of the action space, like in locomotion and the cases of manipulation that poses the most issues. The experiment use classic but simple 2D navigation and simulated locomotion scenarios to illustrates the benefit.

Weaknesses

The comparison is rather limited, we woud have like to see DADS, MUSIC and LSD for example. We would also have like to see experiments in more known challenging environment like manipulation where MI is the most in trouble.

Questions

Why not having compare with DADS, MUSIC and LSD? Is the random walk used at the last step of the guide skill discovery is scalable to larger state environments ?

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

1 poor

Limitations

The experiments are rather limited to justify the benefit of the approach. The random walk as last step of the definition of the guide skill definition isn't shown to be scalable to larger state environments.

Authorsrebuttal2023-08-16

Kind reminder for discussion

Dear Reviewer gakY (reviewer 1), We deeply appreciate the time you've dedicated to reviewing our paper. While we believe we've addressed the concerns you raised, we have yet to receive your feedback. We feel that further discussion and feedback would greatly enhance the quality of our paper. We kindly request your continued engagement and insights. Best, Paper 12428 authors

Area Chair q19X2023-08-18

Dear reviewer gakY, you raised concerns about comparisons with baselines DADS, MUSIC and LSD. The authors have responded to this first by explaining their choice of comparisons, and also by providing experiments with LSD. May I ask you to acknowledge the rebuttal and to tell us (and the authors) what you think? Thanks, AC

Authorsrebuttal2023-08-18

Dear Reviewer gakY, Regarding the comparison with LSD, in addition to our prior LSD experiments on 2D bottleneck maze, we are in the process of organizing an extra experiment on the Ant-$\Pi$ maze to demonstrate the robustness of $r_\text{exploration}^\text{DISCO-DANCE}$. We hope this additional experiment will address the Reviewer gakY's concerns well. We intend to share the outcomes by Aug 20th 2am EDT. Best, Paper 12428 authors

Authorsrebuttal2023-08-20

Comparative Analysis of DISCO-DANCE and LSD

Dear Reviewer gakY, We appreciate your feedback regarding the need for comparisons with algorithms such as DADS, MUSIC, and LSD. Initially, as discussed in our response, we classified skill discovery methods into two broad categories: 1. Algorithms that utilize $r_{skill}$ with a mutual information objective while focusing on constructing an effective auxiliary exploration reward $r_{exploration}$. 2. Methods that seek to redefine $r_{skill}$ beyond mutual information. In our earlier communication, we positioned DISCO-DANCE in the former category, believing that its closest counterparts were the algorithms targeting the exploration reward. This informed our initial decision to exclude DADS, MUSIC, and LSD from our comparisons, as they were primarily aligned with the latter category. However, upon your feedback and deeper reflection, we recognize the value of broadening our comparative analysis. Even if DADS, MUSIC, and LSD primarily fall into the second category, these algorithms may overcome the inherent pessimism associated with the mutual information objective. In response, we have decided to extend our comparative analysis with LSD to ensure a comprehensive evaluation of skill discovery algorithms. Among those, we selected LSD, the algorithm known as the top performer among the reviewer has suggested. Due to time constraints, we narrowed our focus to the Ant-$\Pi$ maze, the environment where the agent suffer from exhaustive exploration. | $r_\text{skill}$ | $r_\text{exploration}$ |state coverage | | -------- | -------- | -------- | | DIAYN | None | 22.50±3.34 | | DIAYN | DISCO-DANCE | 39.00±4.85 | | LSD | None | 38.80±3.34 | | LSD | DISCO-DANCE | 45.80±3.34 | The above table illustrates the performance comparison with LSD. A noteworthy observation is that, unlike in the 2D-bottleneck-maze (Fig.c in attached pdf), both DISCO-DANCE and LSD exhibit comparable performance in the Ant-$\Pi$ maze. This potentially underscores LSD's capability to mitigate the inherent pessimism of mutual information objective. While the performance of LSD is noteworthy, it doesn't diminish the significance of DISCO-DANCE. This is because $r_{skill}$, when decoupled from the mutual information objective, and $r_{exploration}$ as guided by DISCO-DANCE, can work in tandem, complement each other. We tested this by combining DISCO-DANCE with LSD, which led to significant performance improvements. Thus, in environments that are challenging to explore, DISCO-DANCE can serve as a key role in improving exploration. In summary, we acknowledge that our initial experiments primarily utilized MI-based algorithms, specifically DIAYN. Yet, as shown in our additional experiments, DISCO-DANCE not only outperforms LSD in both the 2D-bottleneck-maze and Ant-$\Pi$ maze, but also exhibits augmented performance improvements when integrated with LSD. We are also in the process of conducting experiments with MUSIC to provide a more comprehensive evaluation across various environments. In our revised manuscript, we will incorporate these results, along with additional ablation studies where DISCO-DANCE is utilized on top of different skill discovery objectives. We hope our response alleviates your concerns. Best, Paper 12428 authors

Authorsrebuttal2023-08-17

Thank you for the constructive feedback and positive evaluation. We will provide more details about UPSIDE in the Appendix.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC