__ Summary and Contributions__: *** Update: I have read the author response, which was very nice. I maintain a high opinion of this paper, and think it should be accepted. ***
This paper proposes a Bayesian model to adapt to nonstationary dynamics in RL problems. The paper draws on DPs, GPs, an other ideas to create a principled model that is robust and data efficient. A small but well-done experimental section compares against several natural baselines, and demonstrates the superiority of the approach.

__ Strengths__: I liked this paper. This model feels principled, clearly defined, and technically sound. I felt that the paper was well-written, with many design choices that seemed solid. The posterior inference algorithm is complex, but not unnecessarily so.
The experiments were small, but reasonably well done. I especially appreciated the inclusion of several natural baselines. The method seems to work well, and does the job it was designed to do.

__ Weaknesses__: The model is fairly complex, and some of the design choices seem a bit ad hoc. As this is a Bayesian model, there ought to be principled posterior distributions for every quantity; your "algorithm" should then be inference in those posterior distributions.
For example, you have somewhat hack-ish methods of splitting and merging experts, but this could be done using appropriate MCMC proposals, such as birth and death kernels. Likewise, the argmax assignments feel like an approximation to a true posterior.
In general, I think the paper could be strengthened by clearly laying out the normative quantities, and then discussing how your algorithmic choices approximate those, and what the tradeoffs of such approximations are likely to be.

__ Correctness__: Yes

__ Clarity__: Yes - I found this to be a very well-written paper.

__ Relation to Prior Work__: Yes

__ Reproducibility__: Yes

__ Additional Feedback__: Well done!

__ Summary and Contributions__: The paper proposes a continual online model-based RL approach that does not require pre-training to solve task-agnostic problems with unknown task boundaries. In order to achieve this, the authors mainly maintain an infinite mixture of experts and represent each different type of dynamics with a GP (Gaussian process). With the dynamic models, it uses MPC to make decisions. And to adapt the changes of dynamics, it uses a transition prior and proposes the expert merge and prune mechanism.

__ Strengths__: The framework proposed by the authors can handle online nonstationary tasks without requiring task delineation or depending on a pre-trained, which is not able to be realized in meta-learning framework. Also, it is closer to the practical scenarios in application. Since it utilizes GP rather than DNN as the dynamic model, it is more explicable, and the time complexity is estimable.
According to the experiments shown in the paper, the framework outperforms other algorithms like MAML in Cartpole, HalfCheetah and Intersection in general.
Nowadays, DNN is widely used in different fields and in various tasks, I think this work will give us some inspiration and look back to some classic methods.

__ Weaknesses__: I think more detailed derivation process should be included in the paper. Additionally, the environments tested in the paper seems not that complex, I mean, it will be better if experiments in environments with higher dimensions can be performed. As it’s declared in the paper, the performance of model-based methods heavily depends on the accuracy of the learned dynamic models. So in higher dimension environments, does the infinite mixture of GP still outperform the previous dynamic models like DNN?

__ Correctness__: As far as I concerned, I don’t think there exists claims or methods incorrect in the paper. Also, the empirical methodology is reasonable.

__ Clarity__: From my point of view, the paper is well written. The authors firstly introduce the background knowledge of continual learning and model-based RL and point out the problems remaining in the field of learning in online nonstationary tasks. And then the authors represent their framework from infinite mixture of GP to expert merge and prune mechanism. Finally, it compares their framework with five model-based baselines in experiment part with understandable charts and shows the strength of their algorithm.
But it’s noted that the derivation process of some equations like equation 2, in my opinion, should at least be included in the supplemental material rather than just mentioned in the references.

__ Relation to Prior Work__: The authors mainly compare their work with prior work in the field of meta-learning and continual learning and clearly explain the difference of this work with the previous work.

__ Reproducibility__: Yes

__ Additional Feedback__: As mentioned above, the paper will be better if more complex environment and more detailed derivation processes can be included.

__ Summary and Contributions__: This paper proposes a method for continuous learning in reinforcement learning where tasks regularly change, tasks remain available for some time and can be recurring but changepoints are not provided to the algorithm. The proposed approach is model based and the model of the dynamics is an infinite mixture of Gaussian processes. The model is compared to several baselines on Cartpole-SwingUp, HalfCheetah and Highway- Intersection where tasks are created by changing the dynamics of the environment. Experiments evaluate change point detection and new/previous task identification, task performance and data efficiency with respect to baselines.

__ Strengths__: The paper is relatively easy to read.
Experiments are well described, show the advantages of the proposed model, baselines are well chosen and ablation studies are also performed.
I like that the approach is not based on deep learning, it is good to have some variety.
The problem addressed is important.

__ Weaknesses__: The impact of hyperparameters or good values are not discussed, unless I missed it.
The evolution of the number of models in the mixture is not discussed. See Additional feedback.
The experiments are performed on simple environments, so it is not clear whether the model works in complex environments.
The idea is original as far as I can tell but the paper relies on several previous works for steps of the learning algorithm which sometimes make it hard to understand the choices made.

__ Correctness__: The paper appears to be correct although I wonder whether there is a typo in equation 3. I was also not convinced of the superior data efficiency of the proposed approach. In my opinion this is not well established by the experiments, although I want to stress that superiority was established on other criteria and I do not think this is a major issue. More details below.
In equation 3, beta is described as a sticky parameter that increases the probability of self-transition. However, it seems to be added to all sums regardless of the value of successive states z. Could there be a typo in the equation, for example a missing rho_{n-1}(z_ik), or am I missing something?
In the experiments, I would suggest comparing the proposed method to DP mixture of DNNs with the transition prior, to better establish whether the GP or the transition (or both) are responsible for the improved performance. With the current experiments, I think it is not possible to claim that the proposed method is more data efficient than [8].

__ Clarity__: The paper is overal clear and well written. I have a few suggestions to make it even easier to understand and/or fix some minor inconsistency. There is no need for the authors to answer to these points as I think the paper is already rather clear.
I am unsure what Figure 1 represents. I might have missed it, but I think pi is not defined. Furthermore, for me, the graphical model representation seems to match the flow of the online learning algorithm rather than the generative distribution. To give some examples, in Figure 1a, model g_t depends on g_{t-1} and x_{t-1}. Furthermore, the notation in figure 1b might not be coherent with the text of section 2: I suspect that the label of the x-axis should be x_tilde, so (x,u) rather than x.
In the definition q^{pr} in line 168, it was not clear to me what the subscript of the sum mean. Is it any variable in the range? Is it any subsequence? Any possible realization of these variables (what I assumed) ?
In equation (4), I understand that z_t is the value assigned previously in the algorithm. I would suggest to use a different notation for this value than for the variable in equation (2,3) which represents the variable.

__ Relation to Prior Work__: I was not familiar with all related works. In my opinion relationships to previous work and differences are well described. My only concern here is that [8,9] are said not to work in practice based on a publication [23] that was 10 years before [8,9]. That being said [8] is a baseline in the experiment so it is not a big concern.

__ Reproducibility__: Yes

__ Additional Feedback__: Models will never be re-evaluated for merge after being seen a sufficient amount of time. Couldn't this lead to an explosion in the number of models? Is there any theoretical guarantee on the rate of increase of the number of models?
possible typo:
l 94: from --> form
----------------------
Dear authors, thank you very muck for the detailed feedback. In particular I am happy to see the additional experiment I suggested.
I have updated my score to a 7, as I feel the paper is worth accepting with the modifications you propose. It might also be interesting to mention mixture compression and the points discussed with R2 as potential improvements.