NeurIPS 2020

Cooperative Heterogeneous Deep Reinforcement Learning

Review 1

Summary and Contributions: --- UPDATE: The authors resolved / explained most of my concerns, some of which were caused by my not-complete undertanding of the presented concepts. The experiments seem to be reported fairly, wrt. the time-steps used. There seems to be no prior work on ensemble of heterogenous RL-agents, meaning this work has a potential to inspire a new direction of research. I still think that the paper would benefit from proof-reading and some more "polishing". I suppose the author will revise the text itself with some of the explanations. With that in mind, I feel that the paper is of interest to the community, sufficiently novel and can be accepted (reflected in the score change). --- The paper presents a framework to combine several distinct RL algorithms together for better learning. A single "global" off-policy agent learn from an experience buffer which is updated with other "local" on-policy agents, which are periodically updated to match the global agent's policy.

Strengths: The idea of combination of several algorithms together to extract their strengths is interesting.

Weaknesses: The paper lack some important details. The exact mechanic of the policy transfer between different algorithm is not given. Given the content, I may assume that "transfer" means a simple copying of the parameters, but I remain unsure. When augmenting the experience buffer with other algorithm, it would be nice to clarify why it does (not) introduce any bias in the data. It seems that the different parts of the framework could be replaced by a different way of "tinkering" with a algorithm or its hyperparameters. E.g., the auxiliary on-policy algorithms are here mainly for exploration, but the exploration of the main off-policy algorithm itself can be easily controlled and I suspect it can, with the right setting, work as good as the given complicated framework. The global and local experience buffer seems more like a hack. Again, with the original algorithm, the size of the experience replay buffer can be easily controlled. These settings are not explored, and it is unclear whether they can or cannot substitute the given framework.

Correctness: The experiment setup does not describe the exact way of reporting the results. The X-axis on the graphs is labeled "time-steps", which I am unsure what it means in this context. The proposed framework consists of 3 algorithms trained together, which generate 3-times more "time-steps", which I am unsure if they are accounted for. The source code is not given to verify.

Clarity: The paper would benefit greatly from proof-reading, or at least using some typo-checking service. It contains a lot of errors, which impede the understanding of the content. The method of tranfering the policies is not explained at all. There's a section called "How to Transfer" which do not give any details on how to transfer. In Preliminaries, the authors give a summary of the common algorithm too concisely. More explanation would help the reader. The first sentence of the abstract is utterly superfluous. Overall, the paper need some more work on its style.

Relation to Prior Work: There is a general related work of RL. No work combining different algorithms is presented (I do not know any, but suspect that there may be. If not, state it clearly).

Reproducibility: No

Additional Feedback: "Cooperative" in the title suggest a multi-agent RL setting, which may confuse the reader. Consider a different name. It would be great to include some description in the figures' captions. E.g. Figure 3d looks weird, and the description why is given in a block of text, hardly to find. I am unsure how the "Broader Impact" section discussion is related to the presented framework.

Review 2

Summary and Contributions: This paper proposed the CHDRL framework to achieve high sample efficiency and maintain good stability by cooperative learning of a group of heterogeneous agents (off-policy, on-policy, and population-based EAs). The three main mechanisms behind CHDRL, i.e., cooperative exploration (CE), local-global memory relay (LGM) and distinctive update (DU) are well illustrated and supported by an instantiation (CSPC) using a group of state-of-the-art off-policy, on-policy, and population-based EAs agents. Experiment results on a range of continuous control tasks show that CSPC achieves higher performance than each state-of-the-art agent from the aforementioned group, respectively. An ablation study on three main mechanisms of CHDRL is also provided. =========================== Update after rebuttal: I still have concerns about the practical usage of the proposed framework in a wider range of tasks, mainly on the choice of extra hyperparameters. I would like to keep my original score.

Strengths: 1) Combining the advantage of off-policy, on-policy, and population-based EAs RL algorithms is an interesting and important topic. 2) The CHDRL framework is technically sound and well-illustrated. The proposed framework appears to be novel.

Weaknesses: 1) The performance of the instantiation of CHDRL framework, i.e., CSPC, heavily relies on the choice of hyperparameters, including that of each heterogeneous agent. However, the paper did not provide enough information on how to choose them. This could limit the practical usage of CHDRL. 2) It would be better to include theoretical analysis, e.g, the convergence guarantee, since the main contribution is a general framework.

Correctness: No explicit incorrect statements.

Clarity: The paper is written clearly.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback:

Review 3

Summary and Contributions: This paper presents a Cooperative Heterogeneous Deep RL(CHDRL) framework to learn a policy by incorporating the benefits from different agents with off-policy, on-policy and evolutionary policy learning. The global agent uses off-policy learning and transfers the knowledge to local agents. While local agents adopting on-policy or evolutionary policy learning to explore the local area. Experiments on MujuCo environment demonstrates the CHDRL is able to achieve the better performance.

Strengths: 1. Cooperative Learning It is interesting to introduce the framework of cooperative learning by incorporating benefits from global and local agents with different policy learning methods to transfer knowledge and trade-off exploitation and exploration. 2. Local-Global Memory Reply The global memory is served to the exploration experience for all agents, while local memory for the simulation of on-policy data. This memory replay is important for the global agent to keep on learning efficiently.

Weaknesses: Possible Computation Cost As we need to learn the policy with more than one agent based on different off-policy, on-policy and evolutionary strategies, the computation cost might be larger than a single agent.

Correctness: The CHDRL algorithm is the combination of different existing policy learning methods. It is an empirical algorithm without any theoretical proof. However, the experiments show that the algorithm is able to achieve reasonable good empirical results.

Clarity: Yes, the paper is clearly and well written. It is not difficult to understand and follow it.

Relation to Prior Work: Yes. Compared to guided policy search or evolutionary RL, CHDRL can benefit not just from off-policy and evolutionary learning, but also on-policy learning schemes. Unlike A3C and IMPALA, CHDRL has global and local agents, and mainly focuses on the cooperation of diverse learning schemes.

Reproducibility: No

Additional Feedback: I have read all the reviews and rebuttal, the authors have addressed my concern. Yes, I keep the score 6.

Review 4

Summary and Contributions: The paper presents a new deep RL framework that combines on-policy off-policy and Evolutionary Algorithms by using 3 key mechanisms, i.e., cooperation exploration, local-global memory and distinctive update. As a sample algorithm, the authors combine SAC, PPO and CEM. The authors evaluate their method on 2 Mujoco tasks and conduct ablation studies. The results are comparable with other methods. In my opininon, the strongest result of the method is that, it performs good on all tasks. While different algorithms fail at different tasks, the provided algorithm benefits from using all 3 methods and converges to good policies in all tested problems.

Strengths: In my opinion the strongest part of the paper is the ability to combine 3 types of algorithms, and the framework is method-agnostic. Although the authors combine SAC, PPO and CEM, it's possible to replace them with others. Another strength is finding solutions in all the problems where some of the algorithms fail to converge on particular problems. In a sense, the framework benefits from all 3 types of learning, in one type does not work, it can still learn.

Weaknesses: The results on mujoco tasks are not that impressive compared to state of the art. They are equivalent or pretty close to other algorithms' performances. IN addition, the proposed framework introduces many new hyperparameters. Although these hyperparameters can provide better results, they require further optimization by the user. Sometimes simpler algorithms are more preferrable in this context.

Correctness: The method and the methodology seems to be correct. I couldn't understand if the figure 3d is missing data or not.

Clarity: Paper is clearly and well written. The authors give detailed intro to different RL algorithms and latest research related to their method. The algorithm is explain well as well.

Relation to Prior Work: The paper clearly discusses the difference of this work compared to previous works. They also provide detailed evaluation on comparison of these methods and ablation study.

Reproducibility: Yes

Additional Feedback: -