Summary and Contributions: This is a paper where they basically try to discover new RL algorithms. The approach is simple: they use a gradient-based direct policy search technique to solve a Bayesian type of RL problem with a rather rich hypothesis space where every element corresponds to a "classical vector of parameters" plus what we could call a RL algorithm.
Strengths: They address a very relevant problem, namely the automatic discovery of RL algorithms. But such an approach for learning RL algorithms has already been used before. The main contribution of this paper would be the way they define the candidate space for the RL algorithms, but this is never well explained in the paper.
Weaknesses: They did not do a good review of the related work. Problem badly formalized. Main contribution - the way they define the set of RL algorithms - not put forward in a proper way. Simulation results carry out on fairly simple problems and not that convincing.
Clarity: Not that well because they do not position clearly their contribution.
Relation to Prior Work: Not well done. See section additional feedback and comments.
Additional Feedback: Page 2: In your related work, you have missed several important works, such as for example those of Francis Maes where he proposes approaches for learning fundamental learning rules for RL algorithms (especially for playing bandit problems), see https://scholar.google.be/citations?hl=fr&user=h8kelPwAAAAJ His approach is very close to yours (same type of objective function). Page 3: The finding of an optimal update policy is in some sense expressed as a Bayesian RL problem (you know a probability distribution over environments as prior) but you never make the connection with this field of research. That’s a pitty. In the work of Maes, it is somehow formalized as such. Page 4. You approach can be considered as a gradient-based direct policy search approach for which you have as evaluation metric formula (1), as search space \eta \times \theta and as optimization method a gradient-based method. The main contribution of this paper is how to define the candidate space of your \neta, something you never define very well. That’s a pitty.
Summary and Contributions: **Update : No new experiments with continuous control have been added, and so I am not increasing my score. Proposes an algorithm that learns an entire update rule for RL (without relying on value functions like prior work). The update rule is parameterized via an LSTM, which produces updates for the policy weights and a semantic vector (which is shown to converge to a notion of the value function), for N timesteps in a lifetime. Meta-learning is done by considering the performance of a population of agents each with its own policy weights, semantic vector and in separate environments, but sharing the same update rule. Authors show that the RL procedure obtained after meta-training on a set of grid worlds can beat human level performance on 14 atari games.
Strengths: 1) The presented approach is able to learn an entire RL update rule. The only prior meta-RL work that can learn an RL rule effective for an entirely different environment is MetaGenRL , but it still uses TD learning for value functions (only the policy update is learned). The authors of this work empirically show that their completely learned update gives better results. Further, there is a thorough analysis of the predicted semantic vector which shows that it does seem to correspond to the notion of a value function and the learned algorithm does seem to be implementing a form of bootstrapping. 2) Impressive empirical result of learning an RL rule from grid worlds that is effective on atari games. The plot of performance against the types of environments meta-trained seems to indicate that such methods can keep improving the learned RL algorithm with more data, making this a promising direction for future work.  : Improving Generalization in Meta Reinforcement Learning using Learned Objectives (Kirsh et al.)
Weaknesses: 1) The formulation isn't very novel in the sense that it reuses many aspects first introduced in prior work : it combines the popolution setup of MetaGenRL  for learning a shared update rule, with the LSTM parameterization of the update rule . (it does however enforce a separation between the agent and the algorithm unlike prior works that used LSTMs). 2) While the existing evaluation is quite thorough on the chosen domains, it would be interesting to see how the learned rule performs on continuous control mujoco tasks/other robotic domains commonly tested in meta-RL, but this is not included. This also requires dealing with continuous state and action spaces, which would be good to address.  : Improving Generalization in Meta Reinforcement Learning using Learned Objectives (Kirsh et al.)  : RL2: Fast Reinforcement Learning via Slow Reinforcement Learning (Duan et al.)
Correctness: The claims and the method seem correct.
Clarity: The paper is well-motivated, clear and easy to follow.
Relation to Prior Work: The paper goes over previous meta-RL work quite well, categorizing methods based on how well the learned algorithm can generalize.
Summary and Contributions: The authors introduce an approach for learning RL algorithms in which both the policy and prediction (analogous to the value function) are both updated by a meta-learned network. After training their approach on simple grid world environments the authors show the algorithm can generalize to Atari, where it achieves a competitive performance with A2C.
Strengths: The concept of learning both the policy update and prediction is interesting and a natural progression from previous results. The empirical results are particularly convincing, generalizing from the toy environments to a much more challenging domain, Atari. Furthermore, the approach appears to be surprisingly straightforward/elegant if one ignores all the (presumably) necessary minor details. The analysis and ablation on the method is informative and interesting and the supplementary material contains enough details to reproduce the method even without the code.
Weaknesses: The main distinction of this approach over previous results is the prediction/value function is learned as well as the policy update. The input to LPG is a 5-dim vector (r, terminal, gamma, y(s_t), y(s_t+1)) where the predictions y are reduced to a scalar. My main concern is that the information provided to LPG implicitly forces the prediction to be roughly the value function. Or, in other words, its unclear that LPG might predict anything very different than a function of future rewards. This makes me feel like the main result of the paper is closer to re-discovering RL algorithms rather than discovering new approaches, and the claim that the RL algorithm learns its own prediction is over-stated. The results with LPG-V possibly suggest otherwise but there was a lot of design choices made (regularizers, architecture, hyper-parameter tuning, etc) and unless LPG-V was as carefully optimized, it's not surprising that arbitrarily plugging in a VF learning rule would perform worse.
Correctness: The empirical methodology appears to be correct and useful analysis is provided. Each aspect of the approach seems well-motivated or studied with ablation.
Clarity: The paper is well-written, the approach is clear, and presentation quality is high (figures and paper organization).
Relation to Prior Work: There is a clear distinction from previous results and the approach feels like part of a clear progression in this research area.
Additional Feedback: An interesting analysis would be to show the prediction output of LPG is sufficiently different from the Bellman equation. Additional Questions: Is passing gamma to the LPG necessary since the meta-objective G includes it already? Does LPG still learn well if the input to the LPG is augmented with additional information, such as a condensed representation of the state?