Oracle-Efficient Reinforcement Learning for Max Value Ensembles

Reinforcement learning (RL) in large or infinite state spaces is notoriously challenging, both theoretically (where worst-case sample and computational complexities must scale with state space cardinality) and experimentally (where function approximation and policy gradient techniques often scale poorly and suffer from instability and high variance). One line of research attempting to address these difficulties makes the natural assumption that we are given a collection of heuristic base or $\textit{constituent}$ policies upon which we would like to improve in a scalable manner. In this work we aim to compete with the $\textit{max-following policy}$, which at each state follows the action of whichever constituent policy has the highest value. The max-following policy is always at least as good as the best constituent policy, and may be considerably better. Our main result is an efficient algorithm that learns to compete with the max-following policy, given only access to the constituent policies (but not their value functions). In contrast to prior work in similar settings, our theoretical results require only the minimal assumption of an ERM oracle for value function approximation for the constituent policies (and not the global optimal policy or the max-following policy itself) on samplable distributions. We illustrate our algorithm's experimental effectiveness and behavior on several robotic simulation testbeds.

Paper

Similar papers

Peer review

Reviewer Qu2b6/10 · confidence 3/52024-07-01

Summary

The paper considers the setting in which several good policies are available for some Markov Decision Process, and the agent has to learn how to combine them in a way that allows to achieve higher performance than following a single of the constituent policies. This problem is quite wide and a large number of methods exist to combine policies, with varying assumptions and properties. This paper proposes a method that can realistically be implemented, by considering that per-policy value functions can be learned for an ever-increasing horizon $h$, and then the action to execute in the environment is the one executed by the policy that has the largest value in the current state and current horizon. The novel aspect of the contribution seems to be the reliance on $h$ and the iterative nature of the algorithm. In MDPs with a finite time horizon, we have a finite amount of $h$ values, so a finite amount of value functions to learn. Combined with the fixed amount of constituent policies, this leads to an algorithm whose complexity scales well even to very large state-spaces, because the size of the state space does not intervene in the compute requirements of the algorithm. The proposed method is discussed from a theoretical perspective, and promising empirical results are provided. The experiments consider complicated robotic tasks, with an interesting setting and motivation for this work: the constituent policies are almost-optimal policies for various tasks, and the new policy to learn from a combination of the existing ones is learned against a new task. So, this work seems to be applicable to multi-task RL.

Strengths

The proposed method seems sound, easily to implement (provided that the paper is made clearer) and to lead to impressive empirical results. The assumptions and limitations of the approach are well-discussed, which helps deciding whether it would apply well on some specific problem.

Weaknesses

While the contribution seems of high quality, the paper lacks clarity and intuition, which may make it difficult to reproduce. - Examples of oracles for the value functions should be given, to better indicate to people whether the oracle can be a value function learned on another task, or obtained from rollouts, or requires a simulator. The oracle is an important part of the algorithm, as it is queried $K$ times per horizon step and its output is directly used to perform an argmax operation (the result of the oracle does not seem to be distilled to some learned function) - The paper should be a bit more explicit about how to produce the $\mu_h$ distributions and the fact that this requires resetting the environment and performing actions in it. Not every environment is resettable at will by the agent, and executing all these actions requires an online setting. - The core of the paper is the use of approximate max-following policies, defined in Definition 2.3. The definition is very dry and the reader has to carefully look at the notations to understand where everything comes from. For such an important definition, an intuition and maybe an example would have been very useful. Later in the paper (Figure 2), examples of environments and corresponding max-following policies are used in some argument, but without explaining what a max-following policy is. Thus, the definition is very dry, yet at the core of important arguments in the paper.

Questions

Given the average-low clarity of the paper, I may have mis-understood several parts of the contribution. I expect the authors to disagree with some of the remarks written above. I would welcome to be corrected, and for the authors to take the opportunity to improve the paper given the possible ways it can be mis-understood.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The paper does not seem to have potential negative societal issues, and its scientific limitations are well-discussed.

Reviewer Z6cc5/10 · confidence 3/52024-07-12

Summary

This paper presents an approach to enabling Reinforcement Learning (RL) by improving the process of generating an optimal policy. Their method assumes that there is access to a particular class of Markov Decision Process (MDP). The intuition is having a collection of bases determined as constituent policies. Under some theoretically supported assumptions, the policy learning process they propose guarantees consistently obtaining a policy that is at least as good as the best constituent policy and potentially better. The method relies on an algorithm that generates policies from the max-following class. The authors tested their framework with a practical version of such an algorithm (MaxIteration) in 16 robotic tasks. An essential aspect of their method is that it relies on heuristics to create the collection of constituent policies.

Strengths

This paper presents a structured theory consolidated over the definition of a policy function class called max-following. The authors provided a practical algorithm that can generate policies belonging to that class. They proved their proficiency through robotic tasks, compared to an offline RL method called Implicit Q-Learning (IQL).

Weaknesses

This class falls under a set of observations provided by the authors that are necessary for the theory to hold. Also, the bases of constituent policies rely on heuristics. That could restrict the potential application of their training method. I couldn't find the explanation of Policy 0 and Policy 1 in Figure 3. This figure could be more apparent, as understanding your baselines is essential.

Questions

- What are Policy 0 and Policy 1 in Figure 1? - How do you explain the tasks where all the methodologies had a very low success rate?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The authors addressed their method's main limitation: the need for an oracle for the methodology to hold. They claimed that they will address this concern in the future.

Reviewer BR1J5/10 · confidence 2/52024-07-15

Summary

The paper presents an algorithm called MaxIteration for addressing the challenges of RL in large or infinite state spaces. The core idea is to compete with a max-following policy, which at each state selects the action of the constituent policy with the highest value. The MaxIteration algorithm is efficient, requiring only access to an empirical risk minimization (ERM) oracle for value function approximation of the constituent policies, without needing the value functions themselves.

Strengths

1. The algorithm is computationally efficient, scaling well with large state spaces. 2. The paper provides a solid theoretical foundation with proofs of the algorithm's effectiveness. 3. It improves upon existing policies without needing to explore the entire state space. 4. The algorithm's performance is validated by experiments on robotic simulation tasks.

Weaknesses

1. The algorithm assumes access to an ERM oracle, which might not be practical in all scenarios. 2. While its efficiency in simulation tasks, the algorithm might be complex to implement in real-world systems.

Questions

1. Are there any specific cases where the algorithm's performance might degrade? 2. How does the algorithm deal with non-stationary environments or changing dynamics?

Rating

5

Confidence

2

Soundness

3

Presentation

2

Contribution

3

Limitations

Please see the above comments.

Reviewer Z6cc2024-08-13

Answer to Rebuttal

Dear authors, thanks for your efforts in clarifying my comments. Here are some clarifications to the questions from my side that didn't appear very clear: > We are unsure what "this class falls under a set of observations provided by the authors that are necessary for the theory to hold” means. Would you be willing to clarify a bit more? I think you understood what I meant, and thanks to your answer, I can now identify that the observations are not necessary conditions (assumptions) on which to base your theoretical results. > It is also unclear to us what it means that the bases of our constituent policies rely on heuristics. We would appreciate clarification on this point as well. This part comes from the experimental results that require you to use heuristic-based versions of the algorithm. Thank you for clarifying that this is only related to this part and is not required in the theoretical proof. The confusing part comes from the abstract, where you stated: "One line of research [...] the natural assumption that we are given a collection of heuristic base or constituent policies [...]." I believe this part makes the reader assume that the constituent policies come from heuristics, which is why I commented in the first place. Maybe you can adjust the abstract to clarify the difference. > Regarding Figure 3 Thank you for explaining and considering my feedback to clarify what Policies 0 and 1 mean. I think improving the details you provided about them would help make your experiments more understandable. Your clarification about this was very helpful in understanding the underperforming tasks in my last question. Thanks for taking the time to answer all of them; besides what I stated above, I don't have additional comments.

Authorsrebuttal2024-08-13

Dear reviewer Z6cc, we are grateful for your feedback and will make several changes in the next iteration of the manuscript. * We had hoped to convey that simple heuristic policies can sometimes be useful even when they are not complex, but we see now that there is ambiguity in this statement. We will adjust the abstract as you suggested and point out in the main text that constituent policies can, but must not necessarily, be heuristic. * We will also make changes to the caption and text with respect to the description of the policies 0 and 1 including a paragraph similar to what we provided in the rebuttal. Thank you for engaging in this discussion phase, we greatly appreciate it. We hope you are now more positively disposed to our paper and are happy to discuss further if there are any other points of confusion

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC