Synthetic Experience Replay

A key theme in the past decade has been that when large neural networks and large datasets combine they can produce remarkable results. In deep reinforcement learning (RL), this paradigm is commonly made possible through experience replay, whereby a dataset of past experiences is used to train a policy or value function. However, unlike in supervised or self-supervised learning, an RL agent has to collect its own data, which is often limited. Thus, it is challenging to reap the benefits of deep learning, and even small neural networks can overfit at the start of training. In this work, we leverage the tremendous recent progress in generative modeling and propose Synthetic Experience Replay (SynthER), a diffusion-based approach to flexibly upsample an agent's collected experience. We show that SynthER is an effective method for training RL agents across offline and online settings, in both proprioceptive and pixel-based environments. In offline settings, we observe drastic improvements when upsampling small offline datasets and see that additional synthetic data also allows us to effectively train larger networks. Furthermore, SynthER enables online agents to train with a much higher update-to-data ratio than before, leading to a significant increase in sample efficiency, without any algorithmic changes. We believe that synthetic training data could open the door to realizing the full potential of deep learning for replay-based RL algorithms from limited data. Finally, we open-source our code at https://github.com/conglu1997/SynthER.

Paper

References (87)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer oBNH7/10 · confidence 5/52023-06-18

Summary

Authors propose to use diffusion models for generating new data based on online interactions or offline dataset. Generating new samples allows online and offline algorithms to perform better. Method can be useed with any offline method and any online method which utilize replay buffers.

Strengths

Useful idea which works. Both offline and online RL settings are covered as well as both visual and non-visual tasks. Good ablations and large-scale experiments.

Weaknesses

I would recommend to validate the approach using Antmaze D4RL datasets as this domain is much more challenging than locomotion and maze2d.

Questions

Maybe I missed it, but how would performance change in offline setting if we used original data with the synthetic one? Does it decrease performance? I don't understand the motivation to throw away original data otherwise except demonstrating that diffusion produces good data which could be an ablation.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

I'm don't knowvery much about how complicated it is to train a diffusion model to solve an arbitary task so making it work seems like a potential bottleneck. Probably, it is faster and easier to run an additional hyperparameter search for the RL algorithm than make diffusion work.

Reviewer MwZd7/10 · confidence 4/52023-07-02

Summary

The paper presents Synthetic Experience Replay, a reinforcement learning algorithm employing a generative model based on diffusion to enrich the training dataset of a learning agent. The method is adapted to offline and online RL, and compared to traditional augmentation strategies (e.g., the addition of random noise) in continuous control tasks, from proprioceptive and visual inputs, while applied to different policy optimization algorithms. The paper also presents detailed analyses on the approach, including results on the quality of the generated data and on using larger neural networks.

Strengths

**Originality**: the general direction of using synthetic data for training machine learning algorithms has been vastly explored in the past, with some applications to reinforcement learning. The approach presented in the paper is, thus, not outstandingly novel. However, one of the main points in the paper is that the main bottleneck in previous approaches leveraging synthetic data in the context of RL has been the quality of the generative model. To the best of my knowledge, this is the first use of diffusion models for synthetic data generation. **Quality**: the quality of the work is reasonably good. The experiments are well-chosen and the experimental designs are generally good. **Clarity**: the writing quality is generally very high. The paper is well-written and easy to follow in most parts, with a clear presentation of the experimental results. **Significance**: I believe demonstrating the effectiveness of powerful generative models in generating synthetic data for reinforcement learning algorithms to use is important for the community, given the simplicity and generality of the approach.

Weaknesses

**Major Concerns** - My main concern is on the number of seeds employed for the experiments in the paper. I understand the authors might be subject to computational constraints, but I find 4 repetitions only to be generally not enough to fully trust an individual experiment. I would suggest the authors to run more seeds per experiment and, if possible, use the methods proposed in "Deep Reinforcement Learning at the Edge of the Statistical Precipice" (Agarwal at al., 2021) for aggregation across multiple tasks. - In Section 3.1, the paper says that Tabular VAE and Conditional Tabular GAN have been evaluated using the default hyperparameters proposed in the original paper that applies these generative models to tabular data. However, no dataset from the benchmark employed in the original paper contains robotics interactions similar to the one from D4RL, and this thus begs to question of whether better hyperparameters for such different datasets might exist. I believe the paper would be improved by evaluating those approaches after a careful search for optimal hyperparameters. **Minor Concerns** - Some potentially important references are missing from the paper. In particular, I'm thinking of the recent paper on using synthetic experiences (meta-learned) "Should Models Be Accurate?" (Saleh et al., 2022), and of some recent papers applying, albeit with other goals, diffusion models to reinforcement learning such as "Planning with Diffusion for Flexible Behavior Synthesis" (Janner et al., 2022) and "Is Conditional Generative Modeling all you need for Decision-Making? (Ajay et al., 2022). - The clarity of the paper would benefit from a more explicit description of how the diffusion model is employed for generating the data. One can guess it is naively applied to generating each dimension of each state, action or reward, but being explicit is better than being implicit in this case.

Questions

I ask the author to address the concerns highlighted above in my review.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

4 excellent

Contribution

3 good

Limitations

The computational limitations of the approach are stated in the paper, saying that the approach is computationally better than REDQ. However, the comparison is executed at a replay ratio of 20; it would be desirable to understand what is the actual computational cost of the new data generation and training of the diffusion model even at lower replay ratios.

Reviewer 2GdG7/10 · confidence 4/52023-07-06

Summary

The paper proposes Synthetic Experience Replay (SYNTHER), a novel diffusion-based approach to upsample an agent's collected experience in deep reinforcement learning (RL). The authors demonstrate that SYNTHER is effective for training RL agents in both offline and online settings, and can improve performance for small offline datasets and larger networks. The paper presents results in both state-based and pixel-based environments and provides evidence that synthetic training data can enable more sample-efficient and scalable training for RL agents.

Strengths

1. The paper introduces a novel approach, SYNTHER, which employ generative diffusion model to upsample an agent's collected experience. This method can potentially resolves the challenge of data scarcity in RL. Generally, this method is easy to understand and effective, and may have a broad impact on sample-efficient RL algorithms. 2. The authors provide thorough experiments and evaluations in offline and online settings, demonstrating the effectiveness of SYNTHER in improving performance and sample efficiency. The authors also demonstrated the superiority of the diffusion-based model compared to other generative models such as GAN and VAE. 3. The paper is well-structured, with clear introduction, background and methodology. The author also provide open-source code for reproducibility.

Weaknesses

1. The performance improvement of SYNTHER seems to be very limited under the setting of small networks. 2. The paper didn't demonstrate the advantages of using SYNTHER on existing RL algorithms in both online and pixel based scenarios, nor did it study the impact on model-based RL algorithms in online settings.

Questions

1. Can you provide the experimental results of SYNTHER with a model-based RL method without BC? 2. Can you provide the experimental results of SYNTHER in online and pixel-based setting? 3. SYNTHER can only generate 1-step transition, so n-step bootstrap or GAE cannot be used in imagined data. Did the baselines use n-step bootstrap or GAE value estimation?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

SYNTHER have limited results in pixel-based and online settings. This paper has no negative social impacts.

Reviewer jt3y6/10 · confidence 5/52023-07-06

Summary

Building on the recent success of generative models, the authors propose Synthetic Experience Replay to improve how experience or data is used within reinforcement learning algorithms. While data is normally collected through interaction with an environment, the authors suggest that the experience replay can be augmented with additional data now artificially generated from a diffusion model. The paper considers both the online and offline settings of RL, compares with previous tabular data augmentation methods, and measures performance on pixel-based environments.

Strengths

+ [Quality] The authors conduct an exhaustive set of experiments on standard online and offline RL benchmarks. + [Clarity] The paper is well-motivated, organized, and clearly written. Related works are carefully documented and contributions are clear. + [Significance] The results are quite compelling, especially in the case of upsampling from small datasets. Considering that no algorithmic changes are necessary, many works stand to benefit from this simple, yet effective strategy.

Weaknesses

+ [Originality] Augmenting data in the replay buffer with synthetic transitions is not new. For example, previous works have attempted to learn a dynamics model during the course of standard online RL training to fill up the replay buffer with fake transitions. Most of the novelty in this work comes from 1) using a diffusion model to generate data and 2) demonstrating experimentally that this kind of synthetic data can be useful in a variety of settings.

Questions

-

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

-

Reviewer MwZd2023-08-10

Thank you for your work on improving the paper

Dear authors, Good job! Thank you for running this additional experiments and for your answer. My main concerns were addressed and I believe the rigor of the experimentation has been much improved by the new results. Please make sure to put the rliable plots in the final version of the paper. I will raise my score to 7.

Reviewer oBNH2023-08-11

Thank you very much for answering my questions and running additional experiments. I'm increasing the confidence score.

Reviewer 2GdG2023-08-19

Thanks for your response. However, the last question was not effectively answered. I will keep my score at 7.

Authorsrebuttal2023-08-19

Additional Response

Thank you so much for your response. We apologize that we missed the reviewer's final question, as it was not in the original review. We will now answer this here. **SYNTHER can only generate 1-step transition, so n-step bootstrap or GAE cannot be used in imagined data. Did the baselines use n-step bootstrap or GAE value estimation?** This statement is untrue, SynthER can be extended to $n$-step transitions by expanding the input to the diffusion model. We refer the reviewer to Line 305 of our manuscript which addresses this. Concretely, instead of modeling 1-step transitions by diffusion tuples of the form $(s, a, r, s')$, we could model expanded tuples of the form $(s, a, r, s', a', r', s'')$ and so on. This represents only a modest increase in the dimensionality of the input space and can easily be handled by the diffusion model. This would allow us to use $n$-step bootstrap or GAE methods. The proprioceptive results used 1-step bootstraps. The latent offline visual experiments in fact already use a 3-step bootstrap with a latent that is derived from a 3-step transition. Please let us know if this has answered the reviewer's doubt or if we can elaborate further.

Authorsrebuttal2023-08-19

We would like to thank the reviewer again for their time in reviewing our manuscript. As the discussion period is coming to a close, we were wondering if the reviewer had any further questions or queries that we could address. Thanks again!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC