Bootstrapped Model Predictive Control

5. What are the specific implications of using either Q or V for the training of the algorithm?

We did not simply replace Q with V; rather, we changed the way of value learning, moving from off-policy TD-learning (Q-iteration) to on-policy model-based TD-learning.

We summarize the theoretical facts and experimental results related to value learning as follows, to highlight the specific implications of using Q or V:

Theoretical facts:

(1) When using Q-iteration, which follows the Bellman optimal equation, the algorithm must utilize a Q-network for value learning. This is the standard approach in DDPG-style methods.

(2) When using on-policy model-based TD-learning, the algorithm can employ either a V-network or a Q-network for value estimation.

(3) In Q-iteration, it is necessary to use ensemble Q-networks, and select the minimum Q-value for computing TD-target. This is to avoid overestimation bias, which is critical for Q-based methods. However, this increases the number of parameters of the Q-network. Most Q-based methods, such as DDQN, Rainbow DQN, and TD-MPC2, adopt similar technique.

Experimental results:

(1) When using on-policy model-based TD-learning, there is no significant difference between using a Q-network or a V-network in terms of algorithm performance.

(2) Compared to Q-iteration, on-policy model-based TD-learning improves algorithm performance. See the ablation section of our paper, where variant3 is compared with BMPC.

(3) On-policy model-based TD-learning allows BMPC to avoid using large ensembles of value networks, reducing the number of network parameters.

6. Clarification on the trade-off between large N and compounding errors.

When calculating the TD-target, we use the world model to estimate the value of the current policy online. A longer TD-horizon requires the world model to "imagine" further into the future, which can introduce greater compounding errors from the model.

Based on our early experiments, the world model in TD-MPC2 does not perform well in long-horizon predictions, which is why we opted for a shorter TD-horizon (N=1).

We would like to further clarify this issue with the following points:

(1) The "compounding errors" mentioned in our paper refer to the compounding errors of the world model’s predictions. This is a common term used in the MBRL literature.

(2) It is indeed true that increasing N can reduce bias in V-value estimation, but this is only the case when trained with real data rather than imagined data. In MBRL, model errors often have a more significant impact, which prevents us from setting a long TD-horizon unless we have an accurate model.

(3) In fact, although there is a performance difference between N=1 and N=3 on DMControl, it is not substantial.

7. Detailed explanation of "environment step".

The relationship between environment step, inference step, and action repeat is as follows: During an inference step, the agent selects an action based on its current policy. Action repeat refers to the number of consecutive times the same action is applied in the environment without re-evaluating the policy. Each time the action is applied, it results in an environment step, where the environment progresses and returns an updated state and reward. When the action repeat is greater than 1, multiple environment steps are taken for a single inference step, meaning the agent does not update its action choice until after repeating the same action for the specified number of steps.

The term "environment step" is commonly used in the broader RL literature, including in works like TD-MPC2 and DreamerV3. In this work, all DMControl tasks use an action repeat of 2. This means that the number of environment steps is twice the number of inference steps.

1. Overly engineering-focused modifications and limited research insights.

These improvements seem overly engineering-focused

We respectfully disagree with the idea that our work focuses too much on engineering. Here’s why:

(1) We did not tune any parameters of TD-MPC2. For the TD-MPC2 components, we used the exact same hyperparameters as in the original work. BMPC's hyperparameters were also not extensively tuned, as their values were determined through theoretical analysis and minimal experimentation based on their well-understood nature.

(2) Our approach does not simply replace modules of TD-MPC2 with "better" ones or add tricks. Instead, we addressed a specific issue—performance gap between policies—by proposing a simple solution inspired by expert iteration. This is not a trivial "A+B" combination, as expert iteration is a conceptual approach aimed at policy improvement. Applying it to data-driven MPC method requires careful design choices.

(3) While our method incorporates techniques such as imitation and replanning, which may seem more "engineering-oriented", we believe that the problem itself can indeed be addressed through such practical solutions. Sometimes, simplicity is a strength, and introducing new theories may not be needed.

do not provide significant insights into the research direction.

We believe that the issues we identified and the solutions we proposed offer meaningful research insights.

The strong performance of TD-MPC2 motivates us to explore further ways to leverage MPC's capabilities, such as through a bootstrapping approach. Our proposed expert imitation and lazy reanalyze methods are reflections of these research insights. Additionally, the BMPC architecture can be combined with gradient-based policy optimization and can also be applied in offline RL settings, which are further manifestations of our research insights.

Moreover, the issue we identified in TD-MPC2 could be explored further. We hypothesize that it may be related to the comparative strengths and weaknesses of gradient-based versus sample-based optimization in high-dimensional action spaces, as discussed in our response to reviewer tbYo.

While it's easy to make claims, proving them is hard. So, we've avoided overemphasizing unverified insights and instead focused on solving the identified problem with solid experimental results to show that we've addressed it.

2. Describe how expert policy π evolves during training.

If we understand correctly, you are asking how the expert policy (the MPC policy), evolves and improves during training.

The MPC policy π is derived through MPPI planning, based on the world model, value function, and network policy. The process is described in the "MPC with a policy prior" part in the Background section of our paper.

The performance of the expert policy improves as both the world model and value function become more accurate through learning. The more accurate the world model and value function, the stronger the expert policy becomes.

While the specific mechanism of how MPPI makes decisions is not the primary focus of our work, the details can be found in the original TD-MPC paper, where Algorithm1 explains the MPPI planning process in detail.

3. Small number of seeds in experiments.

We followed the TD-MPC2, which used 3 seeds for experiments. From our results, the 95% CI for most curves are fairly narrow, so we believe that 3 seeds are sufficient to support the experimental findings.

However, since two reviewers are concerned about the number of seeds, we will increase the number of seeds to 5 for all experiments. This may take some time, but we will do our best to upload the revised paper with updated results before the discussion period ends (Nov 26).

Additionally, we have already conducted experiments on more domains. Specifically, we extended our experiments on HumanoidBench to verify whether BMPC's performance holds in controlling more complex embodiments. This benchmark requires the agent to control a Unitree robot, with a large action space (61-dimensional). All BMPC hyperparameters/settings remain the same as those used on DMControl.

The results show that BMPC maintains superior performance on the HumanoidBench locomotion suite. We have uploaded the results here, and we will include these new results in the revised paper.

4. Suggestion to mention related work on MPC research.

We appreciate the suggestion to include more related work on MPC research. We will incorporate the references suggested by reviewer qEhR to improve the related work section. The revised paper, along with other changes, will be uploaded before the discussion period ends (Nov 26).

Dear Reviewer Ayiq, Could you please read the authors' rebuttal and give them feedback at your earliest convenience? Thanks. AC

We thank the reviewer for their positive feedback and for recognizing the strengths of our work. Below, we provide our responses to each of the concerns:

1. Clarification on how the MPC policy is computed without re-running the planning algorithm.

Thank you for pointing this out, and you are correct. In our implementation, we store the action distribution of the MPC policy in the replay buffer, specifically the mean and variance of the Gaussian distribution, rather than the actions themselves. We will make sure to clarify this detail in the revised paper.

2. Suggestion to mention related work on MPC imitation.

We will gladly include and cite the related work you mentioned, as it is indeed highly relevant to our study! Due to the limitations of our perspective, we hadn’t fully summarized all relevant work, and we truly appreciate your valuable input, which will help improve the completeness of our paper. We will incorporate these changes along with other revisions and upload the updated paper before the discussion period ends (Nov 26).

Dear Reviewer qEhR, Could you please read the authors' rebuttal and give them feedback at your earliest convenience? Thanks. AC

5. Describe the three variants in Figure 6(a) in more detail.

Facts:

A network policy can be learned by:

(1a) Max-Q gradient; (1b) Expert imitation;

A network policy is needed for:

(2a) Computing TD-target; (2b) Guiding the MPPI process;

The value function can be learned by:

(3a) Off-policy TD-learning (Q-iteration); (3b) On-policy model-based TD-learning;

Variants:

Variant 1: Learn network policy A through (1a), learn network policy B through (1b); use A for (2a), use B for (2b); learn the value function through (3a).

Variant 2: Learn network policy A through (1a), learn network policy B through (1b); use A for (2a), use both A and B for (2b); learn the value function through (3a).

Variant 3: Learn network policy A through (1b); use A for both (2a) and (2b); learn the value function through (3a).

BMPC: Learn network policy A through (1b); use A for both (2a) and (2b); learn the value function through (3b).

We will add further details of these variants in the appendix. Similarly, we will upload the revised paper with these updates before the discussion period ends (Nov 26).

1. Expert imitation introduces overhead into the training pipeline.

We agree that this is true, and unfortunately, unavoidable. Although we have addressed this issue by lazy reanalyze, which has proven to be remarkably effective. Without lazy reanalyze, the training time for BMPC would be 10 times longer.

There are also potential ways to further improve this, such as further parallelizing the reanalyze process by adopting a distributed architecture similar to EfficientZero. While this approach does not reduce the compute overhead, it can decrease the overall training time by scaling up the computation, indicating that BMPC is scalable and can be easily parallelized.

In fact, the re-planning shouldn't introduce this much overhead. The main issue is that MPPI itself is costly. Specifically, for a single inference, MCTS requires 800 simulations in board games in MuZero, gumbel search only requires 32 simulations in various domains in EfficientZeroV2, while MPPI requires 512 * 6 * 3 = 9216 simulations. We tried reducing the number of simulations, but this inevitably led to a decrease in performance to some extent.

However, from another perspective, although MPPI incurs a high computational cost, this might be the reason why it performs so well in continuous tasks.

2. Small number of seeds in experiments.

We followed the TD-MPC2, which used 3 seeds for experiments. From our results, the 95% CI for most curves are fairly narrow, so we believe that 3 seeds are sufficient to support the experimental findings.

However, since two reviewers are concerned about the number of seeds, we will increase the number of seeds to 5 for all experiments. This may take some time, but we will do our best to upload the revised paper with updated results before the discussion period ends (Nov 26).

3. Experiments on more domains.

We have already conducted experiments on more domains. Specifically, we evaluate BMPC on HumanoidBench to verify whether BMPC's performance holds in controlling more complex embodiments. This benchmark requires the agent to control a Unitree robot, with a large action space (61-dim). All BMPC hyperparameters remain the same as those used on DMControl.

The results show that BMPC maintains superior performance on the HumanoidBench locomotion suite. Although the HumanoidBench already includes results for TD-MPC2, we re-ran the experiments using the latest code (which yielded better results). We have uploaded the results here, and we will include these new results in the revised paper before the discussion period ends (Nov 26).

3. Are there other reasons behind TD-MPC's poor value function?

This is a crucial question! On the surface, we tend to think of it this way:

(1) TD-MPC's network policy is trained similarly to SAC, with only two differences: The training data comes from the MPC policy, which is of higher quality; The inputs are well-represented latent vectors rather than raw observations.

(2) SAC tends to fail on high-dimensional tasks.

(3) Hypothesis: Even with better data and better representations, SAC still fails.

(4) Therefore, TD-MPC's network policy also fails.

As for why SAC fails, one potential reason is that for high action-dim tasks, sample-based optimization (like MPPI) is more robust and suitable than gradient-based optimization (like DDPG-style max-Q gradient).

In fact, we tried to combine both expert imitation and max-Q gradient for policy optimization, and it showed fairly good results. However, we have not yet explored this approach in depth. We believe there are many more things to think about, and more experiments are needed to uncover the underlying reasons behind the phenomenon.

4. What is the tradeoff of pure MPC policy distillation compared to model-free policy optimization?

Despite attempts to find any drawbacks, we haven’t identified any clear disadvantages or tradeoffs of MPC policy distillation compared to gradient-based policy optimization in our experiments.

Besides DMControl and HumanoidBench, we also tested a few tasks from MyoSuite and ManiSkill. Overall, BMPC’s performance didn’t fall behind TD-MPC2 in any of these domains. In certain tasks, BMPC might fall behind a little due to an “inappropriate” network policy std, but we found that this can be fixed by adjusting the task-specific entropy loss coefficient (which we did not do), so it’s not a real tradeoff.

In tasks where the MPC policy is weaker than a SAC-like policy (possibly due to a poorly learned world model or an inappropriate planning horizon), BMPC will probably perform worse than TD-MPC2. That said, we have yet to encounter such a situation.

By the way, we also conducted new experiments on HumanoidBench, which has a larger action space(61-dim), and BMPC's performance remained strong without any changes to the hyperparameters. This might help illustrate that the BMPC approach does not excessively interfere with MPC's exploration. We have uploaded the results here, and we will include these new results in the revised paper before the discussion period ends (Nov 26).

1. Concern on bootstrapping between MPC and policy affecting exploration/exploitation.

Will the added policy hurt the exploration of MPC?

No, overall it does not.

First, we observed that adding the guiding policy improves the performance of MPC compared to not adding it, similar to the findings in the TD-MPC2.

Second, entropy loss in policy training prevented potential harm to MPC exploration. In our early experiments, we found that if the policy std is too small, the guiding trajectories may become too narrow, leading to premature convergence to local optima. However, This issue can be mitigated by simply adding entropy loss to policy learning, as in BMPC.

For clarity, the main reason we add exploration in lazy reanalyze was to prevent the bootstrapping mechanism from causing the network policy’s std to shrink over time. In most tasks, this has no impact on performance. To be specific, we aimed to avoid the following scenario:

(1) The std of the MPC policy is always smaller than that of the network policy, since the samples generated by the network policy are iterated by MPPI multiple times.

(2) The network policy mimics the MPC policy, including its std, causing the network policy’s std to shrink.

(3) Eventually, the network policy’s std converges to the minimum value allowed by the MPC policy, which may not be desirable.

Thus, we enforced a lower bound on the std of the network policy during lazy reanalyze to prevent this from happening. However, this situation rarely occurs in most tasks, and even when it does, it typically doesn’t affect performance negatively. It only impacts hard-exploration tasks with sparse rewards a little.

Could you highlight how exploration is achieved in your model, or is it similar to TD-MPC2?

We use the same MPPI configuration as TD-MPC2, including the exploration part. The MPC exploration setting in TD-MPC2 is quite simple but generalizable, which is why we didn’t make any modifications to it.

Do you have to carefully tune or use a curriculum for the std (sigma) of the MPC?

No, we didn’t need to tune anything. We used the same MPPI hyperparameters and configuration as TD-MPC2, including the std. In fact, we kept all TD-MPC2-related hyperparameters in BMPC unchanged to ensure a fair comparison.

2. Concern on the appropriateness of using DreamerV3 for state-based comparisons.

It is true that DreamerV3 was primarily designed for visual tasks (to my knowledge, it doesn’t support multi-tasking). Nevertheless, it also reports state-based DMControl results in the paper, which is the same as our setting. Furthermore, the official code provides hyperparameters for state-based DMControl, so we did not conduct any additional tuning.

3. What's the Effect of EMA in Eqn (7)?

The main purpose of the EMA in Eqn (7) is to normalize the KL loss used for imitation, thereby balancing the KL loss with the entropy loss. Depending on the task, the value of the KL loss of a Gaussian can vary widely, from 0 to 1000, which introduces instability during training and makes it difficult to apply a fixed entropy loss coefficient across different tasks.

How grounded is this design choice?

The similar design choice has been used in both DreamerV3 and TD-MPC2, where it serves the same purpose of balancing the actor loss and entropy loss. Therefore, it is a well-validated design choice based on extensive experimentation.

Dear Reviewer 4Ngs, Could you please read the authors' rebuttal and give them feedback at your earliest convenience? Thanks. AC