NeurIPS 2020

A Self-Tuning Actor-Critic Algorithm

Review 1

Summary and Contributions: This paper proposes a STAC to self tune all the differentible hyperparameters in loss function, and also introduces a novel leaky V-trace operator. On top of these, it introduces auxiliary loss functions to enable self-tuning of the discount factors, leading to a new agent STACX. The paper conducts extensive experiments for STAC and STACX, in atari domain, DM control suite and real-world RL challenge. It shows that STAC consistently improve performance when increasing the number of self-tuning metaparameters. Also it demonstrates the trend of self-tuned metaparameters and robustness to its hyperparameters.

Strengths: 1. This paper proposes a self-tuning actor critic algorithm, using the metagradient to update all differentiable hyperparameters. It proposes a variant of V-trace operator and auxiliary loss functions to allow more hyperparameters to be differentiable. Also, it provides a theorectical guarantee of contration for the leack V-trace operator. 2. I think the main contribution in this paper is the empirical results of STAC. The proposed method was evaluated on multiple domains and demonstrates its effitiveness and significantly ourperformances baseline methods. This sheds light on the using of metagradients in many other RL topics. I like this paper in general.

Weaknesses: My major concern is the novelty about the self-tuning technique in this paper. The way STAC self-tunes the metaparameters by metagradients is straightforward --- introducing a inner loss that parameterized by metaparameters and differentiate the outer loss w.r.t the metaparameters. This is standard way in meta-learning field. Although STAC proposes a novel V-trace operator and auxiliary loss functions, these two contributions only try to make some hyperparameter differentiable for STAC, which are a little bit adhoc. I'd like to see some deep insights from the authors about the self-tuning (metagradients) itself. Other comments: 1. The paper introduces self-tuning metaparameters in the inner loss function. However, the metaparameters are bounded in (0,1) and multiplied by outer prior hyperparameters (g_v^outer, \gamma) which have been well-tuned by other algorithms. Althougn this prior can guarantee the agent is not misleaded by some very large metaparameters, I think this is somehow restricted since the metaparameters are only learned to self-tune the scale of other well-tuned hyperparameters and it can potentially limit the ability of metaparameters. Besides, there is still some hyperparameters in outer loss function. Can we make them differentiable? 2. The paper introduces auxiliary loss functions to enable the self-tuning of discount factor. Is it necessary at all? Have you tried just differentiating the inner loss function w.r.t. the discount factor? At the moment I am not sure which is a better way. 3. It seems that in Atari domain, STAC only slightly surpasses baselines in the later training stage. But in other domains, STAC achieves much better sample efficiency than baselines. Could you explain it a little bit more? 4. Can the self-tuning idea be adopted to off-policy Q-learning? DQN-based algorithms seem more sensitive to hyperparameters than A2C, such as replay buffer size.

Correctness: Yes.

Clarity: The paper is well written.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: ** After reading response ** I think the paper would be interesting to researchers at NeurlPS since it demonstrates the power of self-tuning and perhaps the idea could be used in a wide range of of RL algorithms.

Review 2

Summary and Contributions: The paper proposes to use meta-gradients to adapt the hyperparameters of a modified IMPALA algorithm. The modifications include a version of V-trace that interpolates between truncated importance sampling and ordinary importance sampling, and additional auxiliary tasks based on learning different discount factors. The paper shows improved performance on the Arcade Learning Environment and DeepMind Control suite.

Strengths: The paper is well-written, and attempts to address a very important research question: how can the huge (and growing) list of hyperparameters in deep RL methods be set in a way that avoids expensive and brittle hyperparameter tuning? For this reason the paper seems very relevant to the NeurIPS community, and despite some issues the empirical results look somewhat promising.

Weaknesses: The proposed modifications seem heuristically motivated (and not theoretically justified), with many design decisions made that were not explained or carefully studied. Why a KL divergence penalty in the outer loop to prevent policy changes? Why leaky v-trace over other policy evaluation algorithms that don't clip importance sampling ratios? Why leaky v-trace in the inner loss but not the outer loss? Why add auxiliary tasks to a paper about using meta-gradients to adapt hyperparameters? Why auxiliary tasks solely built around discount factors? The implied justification for these (and other) design decisions seems to be improved performance on the ALE and DM control suite. However, the results are averaged over only 3 random seeds, which makes it impossible to draw conclusions, and alternatives to the design decisions weren't compared. The paper seems to contain two separate ideas: auxiliary tasks based on different discount rates, and using meta-gradients to adapt hyperparameters. It seems like there's not enough space for an in-depth, careful investigation of both.

Correctness: The empirical methodology is weak: only 3 random seeds, and very few alternative design decisions were compared.

Clarity: The paper was fairly clearly written, and did a good job of summarizing IMPALA and the proposed modifications. The paper was very clear on what was done, but not always clear on why.

Relation to Prior Work: The paper clearly discussed how it differs from the closest prior work of Xu et al. (2018), but does not mention—nor compare against—existing methods to adapt specific hyperparameters like the step size or trace decay rate.

Reproducibility: Yes

Additional Feedback: After reading the other reviews and the author response, I have decided to increase my score. I think the main contribution of demonstrating the effectiveness of self-tuning hyperparameters is probably significant enough and interesting enough to the community to warrant publication, despite the issues raised by the reviewers. The paper states multiple times that STACX "reasons about multiple horizons", but there's no explicit reasoning going on; the agent is simply minimizing auxiliary losses with different discount factors that share a layer of weights. This seems like an unwarranted anthropomorphism, and I'd recommend removing these statements. If the bias introduced by v-trace's clipping of importance weights is undesirable—implied by leaky v-trace partially undoing the clipping—why not avoid the clipping in the first place? Why not consider other solutions to the problem or other off-policy policy evaluation algorithms that don't introduce bias by clipping? Averaging results over 3 seeds is not enough to say anything about statistical significance of results or draw meaningful conclusions. Similarly, plotting half of a standard deviation is misleading: non-overlapping fractions of a standard deviation communicates nothing about statistical significance (although in some experiments the .5 standard deviation actually does overlap), yet gives the illusion of statistical significance to an unwary reader who mistakes them for confidence intervals. It would be better to plot confidence intervals and do enough runs to make them not overlap, which visually communicates to the reader that the difference in performance is statistically significant. If the required number of runs to attain statistical significance can't be done due to computational cost, I would recommend moving to a smaller environment where the proposed method can be studied more thoroughly. After that the methods can be scaled up to demonstrate their effectiveness on the ALE or DM control suite. This has the added benefit that the method being developed won't "overfit" to one environment. The auxiliary tasks seemed to harm performance or not make a difference (the pixel observations results have overlapping .5 std dev regions) on the DM control suite experiments while performing well on the ALE. The discussion section explains this away without actually showing any evidence to support the explanations. Overall, I really like the idea of using meta-gradients to adapt hyperparameters, and I think this method has a lot of promise. It would have been much better in my opinion if the paper had more thoroughly explored hyperparameter adaptation via meta-gradients on a smaller domain, then scaled the ideas up to the ALE and/or DM control suite. It would also be better to separate the auxiliary task idea into another paper; it seems orthogonal to hyperparameter adaptation via meta-gradients, and the space could've been used to more thoroughly investigate the idea of hyperparameter adaptation via meta-gradients. I also appreciate the investigation into hyperparameter robustness in section 5, although again the few random seeds makes it difficult to conclude anything from the results. Overall I think the paper has enough flaws that I'm hesitant to recommend it for publication. However, if the authors sufficiently address the questions and concerns raised in my review (and the other reviews), then I will consider raising my score.

Review 3

Summary and Contributions: The paper proposes a meta-gradient method to self-tune the differentiable parameters of an actor-critic algorithm IMPALA. They also add auxillary loass-functions in order to reason about multiple horizons, useful for learning from limited amount of data.

Strengths: *Tuning hyper-parameters is a tedious task in DRL approaches. While the idea of using meta-gradients is not new, self-tuning hyper-parameters using meta-gradients for IMPALA is a relevant contribution. *The paper implements leaky-V Trace (for STAC) with a theoretical proof of its convergence properties * STAC with auxillary tasks is implemented by introducing additional heads with the need to learn additional meta-parameters. *The experimental evaluation is extensive. Experiments are conducted on Atari as well as DM Control suite and the results verify the performance gains in using self-tuning. Robustness experiments have also been performed.

Weaknesses: *Since the paper applies the existing meta-gradients technique towards IMPALA, the novelty of the work is limited. In fact, Xu et al, 2018 applied meta-gradient to tune lambda and gamma for IMPALA. *Since meta-gradients is applied only towards a specific actor-critic framework (IMPALA), the contributions can be seen as an incremental one. *Fig 4 should be STACX Vs IMPALA, wrong legend? *In Fig 3, were the hyper-parmaters for IMPALA experiements fine tuned? Are they the best values? *The run-time of STAC and STACX is 25% more than IMPALA. How will you justify the trade-off between performance gains and run-time?

Correctness: yes, the claims seem to be correct. Theoretical proofs and empirical evaluations verify the claims.

Clarity: yes, the paper is clearly written

Relation to Prior Work: yes, relevant related literature have been discussed and compared

Reproducibility: Yes

Additional Feedback: