SPO: Sequential Monte Carlo Policy Optimisation

Leveraging planning during learning and decision-making is central to the long-term development of intelligent agents. Recent works have successfully combined tree-based search methods and self-play learning mechanisms to this end. However, these methods typically face scaling challenges due to the sequential nature of their search. While practical engineering solutions can partly overcome this, they often result in a negative impact on performance. In this paper, we introduce SPO: Sequential Monte Carlo Policy Optimisation, a model-based reinforcement learning algorithm grounded within the Expectation Maximisation (EM) framework. We show that SPO provides robust policy improvement and efficient scaling properties. The sample-based search makes it directly applicable to both discrete and continuous action spaces without modifications. We demonstrate statistically significant improvements in performance relative to model-free and model-based baselines across both continuous and discrete environments. Furthermore, the parallel nature of SPO's search enables effective utilisation of hardware accelerators, yielding favourable scaling laws.

Paper

Similar papers

Peer review

Reviewer X2nr5/10 · confidence 3/52024-06-28

Summary

This paper maximizes the expected reward wrt the policy by rewriting it as a log marginal likelihood in a probabilistic model where latents comprise of states, actions, and “optimality” observations. The prior over states is the world model and the prior over actions is the policy which being optimized. The likelihood is the exponentiated reward which can be normalized to make it a distribution over auxiliary “optimality” variables. The marginal likelihood can be optimized via EM where in the E step, we maximize the evidence lower bound (ELBO) in (3) with respect to the variational distribution q and in the M step, we fix q and maximize the objective with respect to the policy π. In the E-step (Sec. 4.2), the paper derives an analytic expression for q via the Lagrange multipliers method (eq. 6). In the M-step (Sec. 4.3), the paper uses this q to maximize the objective with respect to π. The key difficulty is that the q derived in the E-step is only known up to a normalizing constant and so the paper proposes to get approximate samples from q using SMC (Sec. 4.2.1). The experiments compare the proposed method to baselines on several tasks with discrete and continuous action spaces, and show wins. There are ablations showing the tradeoff between using more particles vs unrolling for longer. In addition to performance, the paper argues that SPO is preferable to the popular MCTS method because it parallelizes more easily and performs well across more domains.

Strengths

Combining SMC and RL is interesting and the experiments show strong empirical gains. However, theoretically, I am not fully satisfied with the method itself—or at the very least, I found it difficult to understand the method fully from the description. See questions / comments below. Therefore, I’m giving this paper a reject in its current form but I am willing to raise the score if the method or its description is improved.

Weaknesses

-

Questions

Major - SMC (Sec. 4.2.1): From (6) it seems like there is just one distribution we want to sample from instead of a sequence of them. What’s the sequence of distributions that SMC targets? In other words, in the background section about SMC, it’s unclear what is $p(x)$ and what is $p(x_t | x_{1:t - 1})$; how do terms in (8) relate to this $p$? - If I reverse-engineer a sequence of targets distribution from (8), it should be $p(a_{1:t}, s_{1:t}) = \prod_{j = 1}^t \mathcal T(s_t | …) \pi(a_t | …) \exp(A^{\bar \pi}(a_t, s_t) / \eta^*)$. That is, the final target distribution is the transition times the policy times the exponent of the *sum* of advantage functions. I think this is incorrect since it doesn’t correspond to the sum of per-step rewards. Something that would make more sense to me is to have the incremental weights contain the exponentiated rewards for all time steps instead of the last one, where it’s the exponentiated advantage function. This is obviously a completely different algorithm. - I don’t understand why (5) follows from (4). We have an unconstrained optimization of a regularized objective in (4) which is turned into a constrained optimization problem in (5). I also don’t understand how $\alpha$ turns into $\epsilon$. - Why does (10) rewrite (9) as a constrained objective where it seems like in practice, we’re actually just optimizing (9). - How is the value function and the Q-function learned? These seem critical for the method but I couldn’t find a description. - E-step: I agree that it is difficult to optimize with respect to $q$ since it’s the distribution of the expectation and the term inside the expectation, $Q^q$ depends on it (eq. 4). The paper chooses to fix $Q^q$. I would have expected that it is the $q$ in $\mathbb E_q$ that makes optimization difficult since typically we’d have to resort to the reparameterization trick or REINFORCE. If we fix the $q$ in $\mathbb E_q$, optimizing the expectation is simple: just sample from $q$ and directly optimize $Q^q$. Minor - $\gamma$ in (3) is not introduced before. - Should there be a likelihood term in the incremental weight in (2)? Or is $p(x_t | x_{1:t - 1})$ an unnormalized distribution? $p$ is not defined in this paragraph. - Swapping the prior term $\log p(\theta)$ in (9) to the KL term in (10) is a little odd. Would it simply not work otherwise? - In App. G.1, why is $\mu_q$ dependent on $q$? Other relevant work - [AESMC](https://arxiv.org/abs/1705.10306), [FIVO](https://arxiv.org/abs/1705.09279), [VSMC](https://arxiv.org/abs/1705.11140) - [DVRL](https://arxiv.org/abs/1806.02426) - [SIXO](https://arxiv.org/abs/2206.05952), [NAS-X](https://arxiv.org/abs/2308.14864)

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

-

Reviewer xUvG4/10 · confidence 3/52024-07-11

Summary

This paper propose SPO: Sequential Monte Carlo Policy Optimisation, a RL algorithm with the Expectation Maximisation (EM) framework for MDP. Experiments on both discrete and continuous environments show that SPO outperforms baselines in terms of sample efficiency and scalability.

Strengths

Quality: The paper clearly introduces the background of Sequential decision-making, Control as inference, Expectation Maximisation and Sequential Monte Carlo. Significance: The performance improvement of the proposed algorithm is significant.

Weaknesses

Originality: The paper propose SPO, which is the combination of Sequential Monte Carlo (SMC) sampling with the EM framework. The novelty is limited. Quality: The paper does not discuss what problem the paper aims to tackle. The paper uses SMC to better estimate the target distribution. But the paper does not give any experimental results showing better estimation results. Ablation studies are missing. Theoretical insights of proposition 1 are missing. Clarity: The paper is hard to follow as it is full of formulas.

Questions

See the weakness.

Rating

4

Confidence

3

Soundness

2

Presentation

1

Contribution

2

Limitations

Yes.

Reviewer R8DZ7/10 · confidence 4/52024-07-11

Summary

This paper introduces a novel iterative, particle-based approach to sequential Monte Carlo planning by combining expectation maximization with importance sampling. Their approach uses a model to sample real state-action trajectories in a problem domain then, after importance-weighting said transitions, computes a target policy distribution for the agent. This target policy is used for policy improvement. The paper supplies a theoretical derivation of their approach and provides experimental results against baseline algorithms.

Strengths

**Originality**: The paper introduces a novel combination of previously established algorithms to tackle sequential decision making. **Significance:** The results of the paper demonstrate that their proposed approach is *easily parallelizable*, *per-iteration faster* and *explicitly tackles a shortcoming of a prior approach (i.e., addresses weight degeneracy)*. All these would make their work interesting to the planning community. **Clarity:** The paper is well written and easy to understand. **Quality:** The motivation for the work is strong as they study a problem of general interest to the community. The conceptual rationale for their approach is sound and the theoretical build makes sense. The experimental evaluation focussed on the claims the authors made and back up said claims. They not only show performance improvements against commonly used approaches but also improvements on computational costs.

Weaknesses

I think that an evaluation with an imperfect model should have been included rather than left as future work.

Questions

Line 204: How exactly is $\hat{A}$ computed? It isn't very clear on that. How much does removing samples based on weight affect exploration? Is this only helpful when you have a perfect model?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

They broach the limitations of their work sufficiently.

Authorsrebuttal2024-08-12

Reviewer Engagement

Dear AC/Reviewers: We kindly highlight the feedback window is coming to a close, and would appreciate feedback from the reviewers in response to our clarifications of their reviews. We feel we addressed the raised concerns. Specifically, Reviewer X2nr, we clarified their theoretical concerns and engaged constructively regarding the method description. We had hoped for a further interaction, as per their comment "I’m giving this paper a reject in its current form but I am willing to raise the score if the method or its description is improved.". Thank you for your time.

Reviewer R8DZ2024-08-12

Acknowledgement of the rebuttal

I acknowledge that I have read the rebuttal and other reviews and I maintain my score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC