Monte Carlo Tree Search with Boltzmann Exploration

Monte-Carlo Tree Search (MCTS) methods, such as Upper Confidence Bound applied to Trees (UCT), are instrumental to automated planning techniques. However, UCT can be slow to explore an optimal action when it initially appears inferior to other actions. Maximum ENtropy Tree-Search (MENTS) incorporates the maximum entropy principle into an MCTS approach, utilising Boltzmann policies to sample actions, naturally encouraging more exploration. In this paper, we highlight a major limitation of MENTS: optimal actions for the maximum entropy objective do not necessarily correspond to optimal actions for the original objective. We introduce two algorithms, Boltzmann Tree Search (BTS) and Decaying ENtropy Tree-Search (DENTS), that address these limitations and preserve the benefits of Boltzmann policies, such as allowing actions to be sampled faster by using the Alias method. Our empirical analysis shows that our algorithms show consistent high performance across several benchmark domains, including the game of Go.

Paper

References (37)

Scroll for more · 25 remaining

Similar papers

Peer review

Reviewer f8iz6/10 · confidence 4/52023-06-30

Summary

This paper introduces two variants of MCTS designed to perform more exploration during search than the commonly used UCT search policy, while resolving an issue with a previously introduced MCTS variant, Maximum ENtropy Tree Search (MENTS), which was designed to explore more than UCT, but which does not converge to the optimal policy / value function with an increasing number of MCTS trials. To address this issue, the authors modify MCTS in two different ways. In the first variant, Boltzmann Tree Search (BTS), the authors use a Boltzmann search policy during the selection/search phase of MCTS, instead of using UCT, but perform a regular Bellman backup to compute state and action values. In the second variant, Decaying ENtropy Tree Search (DENTS), they modify MENTs to separately keep track of the policy entropy and the (unmodified) value functions, allowing them to slowly decay the entropy contribution over time, while using the entropy-augmented value functions to drive exploration during search. Both variants of MCTS are consistent -- i.e. they value function they compute for the root node converges to the optimal value as the number of MCTS trials increases, and the simple regret decays exponentially with respect to the number of MCTS trials. The authors evaluate BTS and DENTS against baseline methods on both toy MDP problems, and in a 9x9 Go round-robin tournament. They find that BTS and DENTS perform similarly or slightly better than existing baselines on the toy problems, and consistently outcompete other methods (modified to make use of neural-networks) in the Go tournament.

Strengths

- The paper clearly explains an existing problem with MENTS not converging to the optimal root-node policy with increasing trials (e.g. using the D-chain problem to illustrate the point). - The "fix" that this paper proposes to MENTS is intuitive and clearly explained. - The theoretical results appear to be sound, with extensive proofs (though I did not check these carefully). - The empirical results demonstrate that BTS & DENTS can improve upon baseline search policies like UCT and MENTS in non-trivial problems (i.e. Go), while generally performing no worse than the baselines on simpler problems.

Weaknesses

- The motivation for increasing exploration during planning/search could be better explained. - In particular, I think the authors could better highlight that UCT was justified in the context of selecting actions in the real-world, and that per Tolpin & Shimony et al [1, 2], using UCT in *simulation* doesn't make as much sense, since computational actions are less costly than actual actions. - This better motivates the use of a Boltzmann search policy, since exploring more when "thinking" could ultimately lead to better exploitation when finally acting / executing. - Similarly, the authors could better explain their choice of using simple regret (instead of cumulative regret) as the metric for an MCTS algorithm, following the discussion / reasoning in [1, 2]. Cumulative regret doesn't matter if the cost of computation is effectively zero (and only matters slightly if computation cost is much smaller the actual cost of acting). - I don't think this should count very much against the paper, given the extensive effort at formally proving the stated results (which I think is valuable, and should be recognized), but it's arguably the case that BTS are DENTS are fairly incremental and straightforward extensions of MENTS. [1] David Tolpin and Solomon Shimony. Mcts based on simple regret. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 26, pages 570–576, 2012. [2] Nicholas Hay, Stuart Russell, David Tolpin, and Solomon Eyal Shimony. Selecting computations: Theory and applications. arXiv preprint arXiv:1408.2048, 2014.

Questions

- Why does the search temperature $\alpha$ in DENTS differ from the entropy weighting term $\beta$? By default (e.g. in MENTS) this should be the same -- is there a reason why it makes sense have them vary separately? Why not decay $\alpha$ with increasing node visits as well? - In the AR version of DENTS used for the Go experiments, both $\alpha$ and $\beta$ are decayed. AR-MENTS is then described as a variant of AR-DENTS, where $\alpha$ is set equal to $\beta$. But in AR-MENTS, are $\alpha$ and $\beta$ set to a fixed value, or decayed with increasing node visits? I'm assuming it's the former, in which case that should be clarified in the Appendix, but if it's the latter, how is AR-MENTS different from AR-DENTS? - What is the value of DENTS over BTS, or vice versa? Are their high-level rules for when one should be preferred over the other? Minor comments: - There seems to be some inconsistency in notation in equations 11 and 19 vs. the text surrounding them -- $N(s)$ and $N(s, a)$ are used in the equations, but the text uses $m_s$ and $m_a$. Post-Rebuttal Response: The author's have addressed my questions, and provided additional experiments showing how their proposed methods compare favorably against existing baselines. I continue to recommend acceptance.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

One benefit of computing the max entropy policy, as MENTS does, is that it might help with exploration when *acting* (not just planning) --- this is perhaps helpful when the agent doesn't have an accurate model of the world, and needs to learn it over time. Adding some discussion on this point would be good, since it seems like a potential limitation of Apart from that, the authors have addressed the limitation of BTS and DENTS that the require $O(|\mathcal{A}|)$ backups relative to UCT, which may hinder performance when actions spaces are large.

Reviewer yYGy5/10 · confidence 3/52023-07-03

Summary

This paper proposes two new algorithms, Boltzmann Tree Search (BTS) and Decaying Entropy Tree Search (DENTS), based on the Maximum Entropy Tree Search (MENTS) method. The authors prove that in certain situations, MENTS fails to converge to the optimal policy. They analyze BTS and DENTS using simple regret and show that these algorithms will recommend the optimal actions. Finally, the authors evaluate various MCTS algorithms, where BTS and DENTS significantly outperform other MCTS algorithms.

Strengths

1. The paper is well-structured, with clear explanations of the concepts and methods used. 2. The paper identifies the limitations of MENTS and proposes two new algorithms, BTS and DENTS, to address the issues of MENTS. The authors conduct theoretical analysis and perform experiments to validate their proposed solutions. 3. The experiments show that the proposed algorithms outperform other MCTS algorithms.

Weaknesses

It is not quite convincing for the baseline program used in the Go experiments. The Komi in 9x9 Go is usually set to 7 (a balanced setting for both players). Black will have an advantage when Komi is set to 6.5. In addition, in Table 1, does PUCT have the same performance (strength) as KataGo? Why not compare the AR-* methods to KataGo directly? Besides, it is quite confusing why the author uses 128 CPU threads. To the best of my knowledge, this is not the best machine configuration for running KataGo. Using 128 CPU threads may even hurt the performance of KataGo. Moreover, in line 266, the authors said they added Dirichlet noise at the root node. In AlphaGo Zero, there is no need to add Dirichlet noise during evaluation.

Questions

1. The authors mention that MENTS may struggle to converge under certain specific circumstances by using the D-Chain as an example. However, it is unclear whether this phenomenon generally happens in other environments. 2. The author mentions the difficulty of selecting β. The experiments also show that BTS (DENTS with β=0) performs better than DENTS. Do the authors conduct any analysis or have any suggestions for β? In what scenarios where DENTS will perform better than BTS? 3. The author claims that DENTS will be similar to MENTS when setting β(m) = α, is there any experimental/theoretical explanation for this part? The relationship is a bit unclear from the mathematical formula. 4. How does MCTS operate in a stochastic environment for Sailing problems? 5. Why were RENTS and TENTS only tested on the 6x6 Sailing Problem and not on the 8x12 Frozen Lake? Other minor suggestions: 1. The lines in Figure 2 are overlapping, making it difficult to see clearly. 2. In equation (4), should it be $V_{\text{sft}}^{\pi}(s', t)$ or $V^{\pi}_{\text{sft}}(s', t + 1)$?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors addressed their limitations in section 5.3.

Reviewer gH3c5/10 · confidence 3/52023-07-05

Summary

<< I have read the authors' rebuttal and have raised my score based on the discussion >> The paper addresses challenges in online tree-based search algorithms, such as UCT, which have shortcomings such as being prone to local optima and a slower convergence rate to the optimal action. An alternative approach, Maximum ENtropy Tree Search (MENTS), despite theoretically optimal performance, has practical limitations like sensitivity to the temperature parameter, a tendency to over explore, and increased computational expense. The authors present two new algorithms to address these issues: Boltzmann Tree Search (BTS) and Decaying ENtropy Tree Search (DENTS). BTS employs a Boltzmann search policy, focusing solely on reward maximisation, whereas DENTS introduces entropy backups that can be used with varying weighting in the search. Both algorithms preserve the exploratory benefits of MENTS, and importantly, they consistently converge to the optimal reward maximising policy. Key contributions of the paper include highlighting the misalignment of the entropy objective in MENTS with reward maximisation, introducing BTS and DENTS, providing an analysis of MENTS, BTS, and DENTS through the lens of simple regret, and demonstrating the improved performance of BTS and DENTS over other MCTS algorithms in specific environments and games.

Strengths

Novelty: The presented Boltzmann Tree Search (BTS) algorithm is a variant of the Maximum ENtropy Tree Search (MENTS), substituting the soft update value in equation 13 with the max operator in equation 19. The Decaying ENtropy Tree Search (DENTS) proposes a novel method for decaying exploration to ultimately converge to the optimal solution, an aspect MENTS struggles with due to its focus on maintaining entropy in the search policy. The novelty of this paper's contributions, while discernible, seems to be limited. Clarity: While the paper is generally well-written, comprehension is somewhat impeded by the multitude of mathematical symbols and technical jargon used throughout. An effort to streamline and elucidate the language could significantly enhance readability. Significance: The paper's experimental results highlight DENTS' robustness to the adjustment of the temperature hyperparameter, a critical element in MENTS. Moreover, the simplicity of the proposed changes augments their potential for practical adoption in the field.

Weaknesses

W1: The novelty of the work seems somewhat limited. It remains unclear whether the improvements result from the incorporation of entropy in the case of DENTS or from the use of hard updates instead of soft updates in BTS. W2: The paper emphasizes the sensitivity of MENTS to the temperature hyperparameter. It is true that many online search algorithms rely on hyperparameter tuning for optimal performance. However, it is noteworthy that MENTS outperforms both BTS and DENTS on the 10-chain and gridworld tasks when the temperature hyperparameter is appropriately tuned. Therefore, it could be argued that the only advantage of using BTS or DENTS is their lesser dependency on temperature. Yet, with the correct tuning of this singular parameter, MENTS could potentially offer a superior solution.

Questions

Q1: I found it somewhat surprising that MENTS performed relatively poorly on the game of Go, especially considering its strong performance in other domains. Can you confirm whether the temperature hyperparameter for MENTS was appropriately tuned in this case? Q2: What methodologies were employed by the authors to tune the hyperparameters for different baseline algorithms, including BTS and DENTS? Q3: To better understand the influence of hyperparameters on BTS and DENTS and gauge their robustness, would it be possible for the authors to compare their performance for varying values of the alpha parameter?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors discussed the limitations of their work. No discussion needed regarding potential negative societal impact.

Reviewer nux66/10 · confidence 3/52023-07-07

Summary

The paper presents two MCTS based algorithms called Boltzmann Tree Search (BTS) and Decaying ENtropy Tree-Search (DENTS). These algorithms overcome the limitation of two state-of-the-art algorithms pUCT and MENTS of under and over exploration respectively. Further, the paper provides simple “simple regret” analysis and shows state-of-the-art performance on benchmark tasks.

Strengths

1. The paper is very well structured and easy to follow for the most part. 2. The empirical results are impressive, especially on Go. 3. The algorithms are novel to my knowledge and are inspiring.

Weaknesses

1. The motivation for analyzing simple regret instead of cumulative regret is not clear. This limits the contribution of the proposed approach to some extent. (I will update the score post rebuttal, given the authors' response). 2. The paper puts all the theoretical proof in the appendix. Please provide at least a proof sketch in the main paper. 3. Section 6.1 is not clearly written. What do you mean by “ To utilize ... a terminal/goal state.” (lines 232-234) Minor 1. Graphs in Figure 2 are of poor quality. Please consider using transparent lines or a variety of dashed lines (different patterns of dashes) to represent different algorithms. 2. Please increase the font size of the text in Figure 3. Note: I haven't carefully taken a look at the proofs in the appendix. I might update my score based on correctness.

Questions

1. Line number 18 - What do you mean by “exact methods”? 2. Why H-MCTS [27] (or other algorithms that focus on simple regret) is not used as a baseline? 3. Are there no approaches that study how to decrease alpha in MENTS so that Theorem 2-like statement holds? 4. How to set the value of alpha in BTS? 5. Why are RENTS and TENTS not compared on Frozen Lake?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

Limitations are clearly mentioned in the paper.

Reviewer f8iz2023-08-12

Thanks for the response.

Thank you for responding to my questions, which helped to clarify my understanding of the paper. I have read the rebuttal and will be maintaining my recommendation for acceptance.

Reviewer yYGy2023-08-16

Thanks for your response. I have a few questions and suggestions. It is weird that using Dirichlet noise performs better than without Dirichlet noise during evaluation. As AlphaZero didn’t use Dirichlet noise in the evaluation. > Our work does not intend to be a new state-of-the-art for Go, and KataGo will beat our algorithms given the same time constraints. I understand that you are not trying to beat KataGo. However, it is still important that compare your baseline model (PUCT) to KataGo to verify your methods. In any case, as you used the same neural network as KataGo, your PUCT should at least has a comparable strength to KataGo even without these heuristics.

Authorsrebuttal2023-08-18

Thank you for your suggestions We agree that we should compare our PUCT agent with KataGo to verify its strength, we were in the process of setting up these experiments this week. We have now run our PUCT agent versus Kata Go and limited them to 1600 trials to give an idea of relative strength. On the 9x9 board, the score was 61(PUCT) - 39(KataGo) and on the 19x19 board, the score was 35(PUCT) - 65(KataGo). We were a little surprised to see our agent win in 9x9, but it could be that KataGo's alterations to UCT are more suited to the 19x19 board, or because our implementation runs slower (so when selecting moves the values used in UCB are likely more up to date) We also agree that adding Dirichlet noise seems unnecessary after you pointed it out in your review. While updating the results when we update our paper, we will continue to try with and without the noise to make sure our PUCT agent is as strong as possible. Given the results from our PUCT agent versus KataGo, it could be that the noise is sometimes useful in 9x9 Go with fewer trials but not so much in 19x19 Go

Reviewer yYGy2023-08-21

Good to see these additional experiments. I would suggest including these experiments for clarity of the baseline (PUCT) model. I have adjusted my score accordingly.

Authorsrebuttal2023-08-21

Thank you, we completely agree and plan to add these experiments from our PUCT agent versus KataGo into the paper to give readers and idea of the strength of our PUCT agent

Authorsrebuttal2023-08-17

HMCTS Results

Hi, we've now added HMCTS to our smaller experiments and can include it in our results. For selecting hyperparameters we varied the budget threshold (to switch from sequential halving to UCT) and the UCT bias parameters. For the UCT bias parameter we used searched over the same values as when we ran UCT independently. On the 10-chain HMCTS always recommends taking the reward of 0.9 for all of the values of budget threshold that we tried {1,3,10,30,100,300,1000}, and so it behaved identically to UCT here. This is because it is still using sample averages to make its recommendations, and the average return from the trials that search down the chain is lower than 0.9. On the Frozen Lake HMCTS also performed similarly to UCT and struggled to explore to find better paths to the goal. The only difference was that HMCTS managed to find its best path in fewer trials than UCT. This would likely happen because the use of sequential halving helps in the upper portion of the trees (for example eliminating actions that directly fall down holes), but ultimately it still relies on UCT and struggles with the same problems. In the sailing domain UCT already performs well and there wasn't any noticeable difference between HMCTS and UCT. We tried the values of {10,30,100,...,30000,100000} for the budget threshold in the gridworld domains. In conclusion, HMCTS's behaviour is very similar to UCT's in our experiments.

Reviewer nux62023-08-17

response to the rebuttal

Thanks for an elaborate response with the new results. Your response clears confusion on experimentation. 1) Can you provide a proof sketch during the rebuttal so that I and other reviewers can verify the correctness? 2) Now I am a bit more clear on simple regret vs cumulative regret, but I think you should perhaps explain this with an example, maybe in the appendix.

Authorsrebuttal2023-08-19

> Now I am a bit more clear on simple regret vs cumulative regret, but I think you should perhaps explain this with an example, maybe in the appendix We can also add an example to the appendix. The use of MCTS in robotics would provide a good example, such as using MCTS for path planning [1,2]. If the robot is equipped with a simulated environment to search in, then simple regret is would be a more appropriate metric, and if the robot can only interact in the real world then cumulative regret would be a more appropriate metric. [1] Janson, Lucas, Edward Schmerling, and Marco Pavone. "Monte Carlo motion planning for robot trajectory optimization under uncertainty." Robotics Research: Volume 2. Cham: Springer International Publishing, 2017. 343-361. [2] Dam, Tuan, et al. "Monte-Carlo robot path planning." IEEE Robotics and Automation Letters 7.4 (2022): 11213-11220. > Can you provide a proof sketch during the rebuttal so that I and other reviewers can verify the correctness? Here are some proof sketches for the results in the paper. We've tried to keep them short so that they can feasibly be put into the papers in a similar form. But please let us know if any more details would be helpful. ### Theorem 5.2 Let $0<\pi_\min<\pi_\text{DENTS}(a|s)$ for all $s,a$, where $\pi_\min$ exists because $\exp(x)>0$ for any $x$ and $1/|\mathcal{A}|>0$ in Equation (21) (Lemma D.1). Then for a DENTS process $\Pr(E)\leq C_0\exp(-k_0N(s_t))$ iff $\Pr(E)\leq C_1\exp(-k_1N(s_t,a_t))$ iff $\Pr(E)\leq C_2\exp(-k_2N(s_{t+1}))$ holds for any event $E$ and constants $C_0,C_1,C_2,k_0,k_1,k_2$, which can be shown using $\pi_\min$ with Hoeffding bounds over indicator variables (Lemmas D.4, D.5). We can then show $\Pr\left(\left|\hat{Q}^{N(s_1,a)}(s_1,a) - Q^*(s_1,a)\right| > \varepsilon \right)\leq C_3\exp(-k_3\varepsilon^2n)$ by induction, where the base case of the induction $\hat{V}^{N(s_{H+1})}(s_{H+1}) = V^*(s_{H+1}) = 0$ holds vacuously (Lemmas D.10, D.15 and Theorem D.16). Let $\Delta_\mathcal{M}$ be a small constant (Equation (105)) such that $\forall a\in\mathcal{A}. \left|\hat{Q}^{N(s_1,a)}(s_1,a) - Q^*(s_1,a)\right| \leq \Delta_\mathcal{M}/2 \implies \psi_\text{DENTS}(s_1)=\pi^*(s_1)$. Then setting $\varepsilon=\Delta_\mathcal{M}/2$ gives a bound on $\Pr(\psi_\text{DENTS}(s_1)\neq\pi^*(s_1))$ which can be used in the definition of simple regret to give the result. ### Theorem 5.1 Theorem 5.1 follows immediately from Theorem 5.2, as mathematically BTS is a special case of DENTS with $\beta(m)=0$. ### Theorem 4.2 $\Pr\left(\left|\hat{Q_{\text{sft}}}^{N(s_1,a)}(s_1,a) - {Q_{\text{sft}}}^*(s_1,a)\right| > \varepsilon \right)\leq C_4\exp(-k_4\varepsilon^2n)$ can be shown similarly to Theorem 5.2. The restriction on the value of $\alpha$ is required, as similar to Theorem 5.2, we need to use $\left(\alpha\leq \Delta_\mathcal{M}/3H\log |\mathcal{A}|\right) \land \left(\forall a\in\mathcal{A} \left|\hat{Q_{\text{sft}}}^{N(s_1,a)}(s_1,a) - {Q_{\text{sft}}}^*(s_1,a)\right| \leq \Delta_{\mathcal{M}}/3\right) \implies \psi_\text{MENTS}(s_1)=\pi^*(s_1)$ to bound the probability of $\psi_{\text{MENTS}}(s_1)\neq \pi^*(s_1)$. ### Proposition 4.1 This follows from computing the values of $Q^*_{\text{sft}}(1,a_1)$ and $Q^*_{\text{sft}}(1,a_2)$ in the modified 10-chain from Figure 1 of the paper using a temperature of $\alpha=1$. (Note that here $a_1,a_2$ refers to the two actions that can be taken in the D-Chain, rather than the variables representing the first and second actions taken on a trial. We will update the figure and paper to avoid this naming clash). We can use $\Pr\left(\left|\hat{Q_{\text{sft}}}^{N(s_1,a)}(s_1,a) - {Q_{\text{sft}}}^*(s_1,a)\right| > \varepsilon \right)\leq C_4\exp(-k_4\varepsilon^2n)$ from Theorem 4.2, to show $\Pr(\psi_\text{MENTS}(1) = a_1)\rightarrow 1$ as $n\rightarrow \infty$ below, so gives MENTS a simple regret that stays above 0 as $n$ tends to infinity. N.B. Original response on 19/8, edited on 20/8.

Reviewer gH3c2023-08-19

Further clarifications

Thank you for addressing the points I raised in my review. Your clarifications are appreciated. Nevertheless, I still have some lingering questions about the scenarios in which BTS and MENTS significantly surpass existing methods. Q1. Can you elaborate on the specific properties of problems that make them more amenable to BTS or MENTS? Q2. It would be beneficial to have a concise table comparing the strengths and weaknesses of MENTS, DENTS, and BTS. For instance, you highlighted reward signal density in your rebuttal. This would help in providing a more clear-cut comparison. Q3. Lastly, how would BTS and DENTS stack up against a method that employs Thompson sampling in MCTS, as presented in Bai et.al. [1]? It's noteworthy that Thompson Sampling also boasts superior regret convergence compared to UCT. [1] Bai et.al., Bayesian Mixture Modelling and Inference based Thompson Sampling in Monte-Carlo Tree Search.

Authorsrebuttal2023-08-20

> Can you elaborate on the specific properties of problems that make them more amenable to BTS or MENTS? Two properties of environments that would make them more amenable to using entropy for exploration are large delayed rewards, and, containing trap states (i.e. states where the agent cannot move to another state for the rest of the trial). The Frozen Lake environment demonstrates both of these properties, there is a large reward that the agent only gets when it reaches the goal, and there are holes that the agent falls into. It may also be the case that there are additional properties that we haven’t considered or thought of yet. We also note that the environment that we considered with these properties (Frozen Lake) was also the environment that exhibited the parameter sensitivity in Figure 26 in the appendix. When the reward signal is dense or engineered so that optimal policy can be found without needing additional exploration, then using entropy for additional exploration may cause unnecessary additional exploration, in which case BTS would be more suitable. > It would be beneficial to have a concise table comparing the strengths and weaknesses of MENTS, DENTS, and BTS. For instance, you highlighted reward signal density in your rebuttal. This would help in providing a more clear-cut comparison. We agree that it would help with clarity, here is a table comparing the different algorithms, and we will include this table in the appendix when we revise the paper. | | UCT | MENTS | BTS | DENTS | |---|---|---|---|---| | Is consistent (will converge to a simple regret of 0 as $n \rightarrow \infty$) for any setting of hyperparameters | Yes | No | Yes | Yes | | Can use entropy for exploration (e.g. helpful for sparser rewards) | No | Yes | No | Yes | | Uses Boltzmann policies for exploration (stochastic action selection) | No | Yes | Yes | Yes | | Optimises for cumulative regret / penalises suboptimal actions during planning | Yes | No | No | No | | Optimised (amortised) complexity to run $n$ trials | $O(nHA)$ | $O(n(H+A))$ | $O(n(H\log(A)+A))$ | $O(n(H\log(A)+A))$ | We also note that using Boltzmann policies has multiple other advantages, such as being able to use the alias method for sampling to make action selection take amortised $O(1)$ time, and the stochastic sampling of actions often leads to higher utilisation of threads when running a multi-threaded tree search. With respect to the complexities, since we discovered that we can use the Alias method for sampling, we considered how efficiently backups can be implemented as that was now the bottleneck. The backups from Equations (12), (13), (18), (23) and (24) from the paper can be implemented in $O(1)$ because only one child value changes. Computing the $\max$ backup from equation (19) takes $O(\log(A))$. Additionally, we could improve on the complexity of BTS and DENTS by approximating the $\max(\cdot)$ operation with a $\tau \log \sum \exp (\cdot / \tau)$ with a suitably small $\tau$, which would match the optimised MENTS complexity of $O(n(H+A))$. > Lastly, how would BTS and DENTS stack up against a method that employs Thompson sampling in MCTS, as presented in Bai et.al. [1]? It's noteworthy that Thompson Sampling also boasts superior regret convergence compared to UCT. We are confident that using DNG-MCTS from Bai et.al. would still struggle with the same issues as UCT. Because it is also designed with cumulative regret in mind, similarly to UCT, it moves towards greedy exploiting policies as more trials are run. For example, if we ran DNG-MCTS on the Frozen Lake environment (note that rewards and actions are deterministic), then the posterior distributions will tend towards degenerate distributions and action selection will tend towards greedy action selection. That is, it would not explore sufficiently to find better paths from the starting location to the goal. BTS, DENTS and MENTS outperform UCT on the Frozen Lake environment because they emphasise exploration more, which is necessary to find improved paths to the goal.

Reviewer gH3c2023-08-21

Thank you for addressing my concerns. I believe it would be beneficial to mention the specific properties of environments in the experiment or discussion section that are more suitable for BTS/DENTS. I have adjusted my score in favor of acceptance.

Authorsrebuttal2023-08-21

And thank you very much for the feedback and we will discuss the properties of environments above in the paper when we revise it

Authorsrebuttal2023-08-21

Thank you to the reviewers

We just wanted to write a thank you to all four of our reviewer's, for the detailed reviews and insightful follow up questions, discussions and suggestions. It has been really helpful to get feedback on what needed more clarification and have a discussion with the reviewer's about how these things can be presented more clearly. Our experimental section has also been improved as a result of running additional experiments during this period. So again, thank you for the time taken to help us improve our work, we are very grateful

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC