NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:4081
Title:Explicit Planning for Efficient Exploration in Reinforcement Learning

Reviewer 1

This paper introduces the interesting idea of demand matrices to more efficiently do pure exploration. Demand matrices simply specific the minimum number of times needed to visit every state-action pair. This is then treated as an additional part of the state in an augmented MDP, which can then be solved to derive the optimal exploration strategy to achieve the specified initial demand. While the idea is interesting and solid, there are downsides to the idea itself and some of the analysis in this paper that could be improved upon. There are no theoretical guarantees that using this algorithm with a learned model at the same time will work. It seems possible that if we specify a large enough lower bound on the entries in the demand matrix, then a PAC bound might be able to be proven that will guarantee that eventually everything will be explored properly. It would greatly help the paper if such a theorem was proven. While the concept behind distance traps is sound, the idea that reward traps are bad is not as clear. R-MAX/MBIE-like algorithms explicitly consider the reward function, because they are designed to eventually converge to the optimal policy and automatically trade-off between exploration and exploitation. This algorithm is doing pure exploration, and thus perhaps a better comparison would be the exploration component of the E^3 which had explicit exploration and exploitation components. You could design domains where MBIE can outperform VIEC, for example the Six Arms domain where as soon as MBIE finds the best arm it says there and no longer tries the other arms; VIEC would continue to try other arms and thus be less efficient. It would be good to clearly make this distinction of purpose in the paper. The final large drawback is the computational cost. The augmented MDP's state space is an exponential blowup of the original state space, making the new problem much more difficult to solve. It would be useful, though perhaps very difficult, to try to derive some simpler approximations of this idea. Overall the idea is interesting, but there is a lack of more in-depth analysis. There is no theory on whether learning without a known model works, nor any considering for how to make this practical. There should be more analysis in at least one of these directions. ** Update after author response ** After thinking more, I have increased my score to a 6, because of the potential of future research with the demand matrix. However I would like to better clarify my view of this paper and hopefully the author can better address these points in the final copy. PAC theorem: I fully do not expect the authors to be able to add this during the rebuttal process, and it was a mistake on my part to put this in the improvements as it was impractical. However I hope that in the future this can be one focus. In the author response, they mention that PAC focuses on worst case, and this is indeed exactly why I mentioned it. Right now, without given a model of the environment, this is no guarantee that in the worst case, some variation of VIEC that tries to learn the model simultaneously will not just fail to satisfy the demand matrix i.e. Algorithm 2 has no guarantee that it will not just sometimes fail. While PAC is not a good analysis tool in comparing exploration, it at the very least is a safety net that says that your algorithm will not outright fail. Reward traps: I brought up the SixArms example as where MBIE is faster than VIEC, and the authors have countered with the fact that given an appropriate demand matrix, VIEC is more optimal. Yes, this is true, but whole point of MBIE is that you do not need to know the optimal demand matrix. Without prior knowledge, you do not know that one of the arms is optimal, and hence you have no way of knowing the optimal demand matrix. The alternative is to use a generic demand matrix such as a constant for all entries, and at that point VIEC will be less efficient than MBIE at solving this task. I imagine that figuring out a near optimal demand matrix may be almost as hard as solving the MDP, in which case VIEC will at best be tied with MBIE for SixArms. The reason for bringing up this example is to highlight that whether reward traps are good or bad is completely domain dependent. In some MDPs you want to be attracted to reward because MBIE will learn faster, and in other MDPs the rewards can be traps which will slow MBIE down. Avoiding reward traps is only consistently a bad thing if your goal is pure exploration. But if the goal was pure exploration, then R-MAX/MBIE are the wrong algorithms to compare to, as their goal is not pure exploration. Like I mentioned in the review, perhaps the exploration phase of E^3 would be a better comparison. But an even easier baseline to compare to is R-MAX/MBIE without giving them the reward signal i.e. letting the real reward always be 0. This is a very simple thing to do if you want pure exploration, and this completely avoids reward traps. My point is that either you should mention that reward traps being good or bad are problem dependent if you want to compare RL performance, or if you want to compare pure exploration then you should simply give reward of 0 to R-MAX/MBIE which basically means reward traps are no longer an issue - this should be a simple change as R-MAX/MBIE with a reward of 0 will still suffer from distance traps in the tower MDPs and I believe all your analyses still hold. The main issue for me is still that there are very large theoretical and practical challenges to this approach. The main practical challenge is the computation cost. You mention an approximate algorithm that works well in practice, and so I hope you include that in the final copy, as that would help introduce a method that could potentially be extended to work in practice. The main theoretical challenges of learning a dynamics model simultaneously and where to get a demand matrix in the first place are difficult and should be for future work. That said, the potential of the demand matrix is enough for me to increase my score.

Reviewer 2

Although the estimate of the computational order is limited to some simple problems, the authors successfully demonstrate when the existing approaches such as epsilon-greedy, R-MAX and interval estimation based exploration strategies result in poor performance compared with the proposed algorithm. The study is quite thought provoking and I believe it would stimulate the future research on this area. ------------------------------------------------------------------------- After reading the other reviewers' comment, and the concerns how to determine the demand matrix is fair. SoI will lower my score from 9 to 8.

Reviewer 3

Originality: it is difficult for me to assess the level of novelty of the ideas proposed in the work proposed in this paper. It seems to me that addressing the exploration problem using an additional planning strategy is close to what has been done in the Bayesian RL literature, in particular, Bayesian approaches doing planning in a Belief-augmented MDP. Quality: it is difficult for me to assess the quality of the proposed approach (at least for its exploration performance). I could not see any empirical evaluation, nor clear theoretical statements (section 4. provides some more formal analyses, but without any claims (I mean Lemma, or Theorem) coming with proofs. Clarity: despite the paper lists contributions at the beginning, I personally found the structure of the paper a bit complicated to follow. In particular, I found Section 4. strangely written: it proposes some form evaluations, but without structuring it in a sequence of Lemma and proofs. Sometimes, the paper develops some intuitions starting from imaginary MDPs, for instance "For example, in an MDP with two states {s1, s2} [...] , a demand D[s1, a] > 1 can never be satisfied and will lead to an infinite exploration cost." These types of arguments could be formalised into lemmas or propositions. Significance: the paper may be of interest for researchers interesting in sample efficiency of RL algorithm. However, the paper should emphasize more (and, perhaps, more quantitatively, including empirical evidence) how the proposed algorithm positions itself of the proposed approach regarding other algorithms. *** Post feedback *** I have updated the overall score according to the feedback provided by the authors.