NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:7148
Title:The Option Keyboard: Combining Skills in Reinforcement Learning

Reviewer 1

Post Response update: Thank you for the detailed response. I still believe that a more in depth discussion of the differences or similarities of policy and cumulant based formulations is required to place the paper appropriately in context of prior work. I think the new results presented by the authors in the response partially address my concerns about comparisons with prior work but not fully. I would still like to see comparison against a policy-based method as per the authors' classification. I agree that all methods might have negative transfer but it would be ideal to include a discussion of the conditions under which the methods would show positive or negative transfer (something that the authors do) and to place that in context with other methods at least qualitatively (something that the authors dont). The newer evaluations in the response do satisfy a part of my concerns. I am still not inclined towards a full-fledged accept but do view the paper more favorably. I would recommend the authors to incorporate the changes suggested by the reviewers. Originality: This paper is largely building off of prior research in hierarchical RL and transfer learning. While the methodology is not very distinct from prior work, this paper does bring them together. The core ideas of this paper are shared by many other works transfer learning to guide hiearchical RL ([1],[2],[3]) and parametrizable options [4]. Significance: The paper demonstrates novel ways of combining hierarchical RL and transfer learning, but with so much other work trying to achieve that (see references), but this is not a unique approach towards that, thus it is hard to gauge significance without direct comparisons. Quality and clarity: The paper is technically sound and well structured. I like the idea of using a linear combination of the Q-values (and cumulants) in a generalized policy improvement framework. I also like the clean reformulation of options into generalized policy evaluation and improvement framework through the addition of termination as a primitive action. Major comments: 1) Coverage of options used: As with most approaches that start with limited number of options or have pre-trained options, the quality of the final solution depends very strongly on the options selected for the 'keyboard'. Thus if all the component actions are poor at the 'test' task, the resulting policy would also have no hope of performing well at it. 2) Distinction from entropy-regularized RL: The author's approach might be seen as a special case of entropy regularized RL as T -> 0 [1]. 3) Benefits/drawbacks w.r.t other continuous parametrizations of options: The authors have provided a single way of parametrizing options through linear combination of their Q functions followed by generalized policy evaluation and improvement. However [3] and [4] provide alternative approaches to parametrization of options. How do these methods compare to yours? 4) Clarity on the cumulant used in Foraging domain: The cumulant function used for each of the nutrients as defined in the paper and the supplementary material differ (line 246 in the paper) and (502) in the supplementary. I would request the author to clarify which one it was, and whether the difference was only in the learning curve for the constituent options or was there a difference in the performance of the options keyboard as well. Minor comments: 1) Behavior of Q-learning with simple options: I am curious as to why simple options showed no learning effects in both the scenarios? [1] - Haarnoja, Tuomas, et al. "Composable deep reinforcement learning for robotic manipulation." 2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2018. [2] - Frans, K., Ho, J., Chen, X., Abbeel, P., & Schulman, J. (2017). Meta learning shared hierarchies. arXiv preprint arXiv:1710.09767. [3] - Gupta, Abhishek, et al. "Meta-reinforcement learning of structured exploration strategies." Advances in Neural Information Processing Systems. 2018. [4] - Da Silva, B., Konidaris, G., & Barto, A. (2012). Learning parameterized skills. arXiv preprint arXiv:1206.6398.

Reviewer 2

Post rebuttal update: I thank the authors for taking the time to prepare a detailed response to my questions, suggestions, and criticisms. I believe that they did a good job in addressing my main questions and concerns. I do agree with other reviewers that comparisons with existing (policy-based) techniques could make this a stronger submission. However, I also agree with the authors that both policy- and cumulant-based approaches should have advantages and disadvantages, so it make sense for them to co-exist in the literature. I suggest trying to better situate the proposed method within the existing literature and to discuss possible downsides of this general approach---e.g. the fact that the simple linear combination of cumulants may often be insufficient to properly span the space of desired behaviors. Given the concerns of other reviewers, I have lowered my score from a 9 to an 8. Overall, though, I still believe that this is an important contribution to the field. ---- This paper introduces a method to combine existing options not merging or analyzing them in the space of policies or actions, but by manipulating them in the space of pseudo-rewards, or cumulants. In particular, the authors first show that every deterministic option can be represented as a cumulant function defined over an extended domain that includes an option-termination action; then, they show that novel options can be obtained by linearly combining the cumulants of the existing options. Importantly, the authors also show that if the cumulants for a given set of existing options is known a priori, the policy for the novel option being synthesized can be directly identified without requiring any learning. This is an interesting paper introducing a novel and important contribution to the field. It is very well written and the experiments clearly demonstrate the contribution of the proposed method. I have few comments and questions: 1) I understand that it may be possible to synthesize novel options by linearly combining their corresponding cumulants. However, I believe there must exist some implicit assumption on what type of behaviors that can be practically defined by computing linear combinations of their corresponding pseudo-reward functions. E.g., the authors introduced, as a motivating example, the goal of learning to walk while grasping an object, given separate options for walking and for grasping. Would it be possible, via the proposed technique, to synthesize options whose constituent options may interfere with one another? The idea of linearly merging reward functions seems to ignore the fact that, in some cases, a reward function for achieving a specific combined goal may need to be more sophisticated than a simple linear scaling of how desirable each of the two (possibly conflicting) sub-goals are. Could you please discuss this possible issue (and the possible underlying assumptions of this approach) in more details? 2) in Eq2, what is "t"? This time variable does not appear anywhere in the definition of the quantity being defined---Q^pi_c(s,a); 3) immediately before Eq3, how is Q^max_c computed, exactly? The text says that Q^max_c = max_i Q^{pi_i}_c, so I assume it corresponds to a maximization over each possible policy pi_i. However, that seems to assume that one Q-function Q1 would always be strictly larger than another Q-function Q2 (i.e., so that Q1(s,a)>Q2(s,a) for all s and a). Shouldn't this max operation, however, be defined over the individual Q-values for each (s,a) under each policy pi_i? 4) the discussion (in page 4, in the paragraph immediately after Eq5) for how the initiation set I_e is defined is not immediately clear to me. It seems to assume that the initiation set is formed by *all* states in which the option does not directly terminate; but this shouldn't necessarily be the case: the termination set may be defined as a strict subset of those states. Could you please clarify or justify this statement/definition? 5) the proposed method seems to assume that, in order to combine options, one needs to know (in advance) the Q-function of each option's policy when evaluated under the cumulant function of all other possible options. Is this correct? If so, how scalable is the method, in the sense that it may require each possible behavior to be evaluated a priori under every other possible pseudo-reward function? 6) in the experiments section (section 5.1) the authors say that "We used Algorithm 3 to compute 4 value functions in Q_eps". It is not immediately clear to me why there are 4 value functions in this case, since m=2 and there are 3 types of good. Could you please clarify? 7) it is very hard to read Fig3 in a black-and-white printed version of the paper. I suggest picking a better set of colors for the curves, if possible.

Reviewer 3

This paper presents a hierarchical reinforcement learning (HRL) framework. The proposed HRL is based on the option framework where the option policies are trained with generalized policy improvement (GPI). Based on the property of GPI, a new option can be obtained by combining evaluated options. The higher-level policy that activates the option is learned through a variant of Q-learning. Originality: Composition of policies used in this paper is briefly mentioned in the original paper by Barreto et al. [2017]. In addition, other methods for composing policy is investigated in [Hunt et al., ICML2019] and [Haarnoja, ICML2018]. Thus, I think the originality of this paper is minor. Quality: The proposed method is a simple extension of GPI to the option framework, and theoretical contribution is minor. Although the empirical results show the benefit of the composing option policies, the performance of the proposed method is not compared with existing methods. Significance: The paper presents a method for obtaining a new option by combining learned options. Although this feature may be novel in the option framework, I think that this contribution is incremental since composing a new policy from existing policies have been investigated in recent studies. === comments after the rebuttal === I appreciate the authors' response. I would like to request authors to address the following points when revising the paper. - What is reported in the rebuttal is comparison of the higher-level policy for composing a new option. The result does not mean that the proposed framework outperforms the method in [ Haarnoja et al., 2018] because the options are not trained with soft-Q learning. Authors should explicitly state this point when addressing the new results to the paper. - I also think that the fact that the options need to be pre-trained is a big limitation since many "policy-based" HRL methods can jointly train the higher- and lower-level policies. This property can be critical in performance since the composition would work very well if the option cumulants properly represents the task dynamics, and if not, it works very poorly as pointed out by other reviewers. In the rebuttal, it is mentioned that "OK and player can be learned together (we are currently working on it).", which means that it has not been achieved yet. I would like to ask authors to honestly discuss the difficulty of jointly training OK and players in their framework.