Grounded Answers for Multi-agent Decision-making Problem through Generative World Model

Recent progress in generative models has stimulated significant innovations in many fields, such as image generation and chatbots. Despite their success, these models often produce sketchy and misleading solutions for complex multi-agent decision-making problems because they miss the trial-and-error experience and reasoning as humans. To address this limitation, we explore a paradigm that integrates a language-guided simulator into the multi-agent reinforcement learning pipeline to enhance the generated answer. The simulator is a world model that separately learns dynamics and reward, where the dynamics model comprises an image tokenizer as well as a causal transformer to generate interaction transitions autoregressively, and the reward model is a bidirectional transformer learned by maximizing the likelihood of trajectories in the expert demonstrations under language guidance. Given an image of the current state and the task description, we use the world model to train the joint policy and produce the image sequence as the answer by running the converged policy on the dynamics model. The empirical results demonstrate that this framework can improve the answers for multi-agent decision-making problems by showing superior performance on the training and unseen tasks of the StarCraft Multi-Agent Challenge benchmark. In particular, it can generate consistent interaction sequences and explainable reward functions at interaction states, opening the path for training generative models of the future.

Paper

Similar papers

Peer review

Reviewer yaXk7/10 · confidence 3/52024-07-08

Summary

The paper proposes a new approach to multi-agent decision-making problems by integrating a language-guided simulator into the reinforcement learning pipeline. Learning before Interaction (LBI) uses a world model to generate trial-and-error experiences and improve the answers to complex decision-making problems. The world model consists of a dynamics model and a reward model, which are learned from expert demonstrations and task descriptions. The LBI framework is evaluated on the StarCraft Multi-Agent Challenge benchmark and shows superior performance on both training and unseen tasks, generating consistent and explainable interaction sequences and rewards.

Strengths

The proposed VisionSMAC dataset is a significant contribution, as it provides a large-scale, diverse, and realistic dataset for multi-agent decision-making problems in the StarCraft II environment. The dataset is constructed by collecting offline trajectories using multi-agent exploration methods and then converting the state into images and languages through a parser, making it easy to create new content. LBI is a novel approach that integrates a language-guided simulator into the reinforcement learning pipeline, enabling the generation of trial-and-error experiences and improving the answers to complex decision-making problems. the LBI framework and the VisionSMAC dataset provide a powerful tool for learning and reasoning in complex multi-agent decision-making problems.

Weaknesses

The motivation for mentioning and comparing the answer of GPT-4, at the beginning, seems misleading. I thought that the proposed LBI would benefit the general response of LLMs on complex decision-making. However, it seems that the proposed method heavily relied on the gaming environment, and and result is limited to SMAC benchmark. The dataset collection is a crucial part of developing a successful LBI, especially for the reward prediction part. However, it must be constructed via a world model (1) where the dynamic is explicit (2) simulation speed is fast enough to perform learning and interaction. How the generalizability of the proposed method?

Questions

See weaknesses

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

no negative societal impact

Reviewer 2fu75/10 · confidence 4/52024-07-12

Summary

This paper proposes a method of improving the sample efficiency and performance of multi-agent reinforcement learning (RL) on the StarCraft Multi-Agent Challenge (SMAC) benchmark by learning to simulate the environment. The simulator consists of an image tokenizer, a dynamics model, and a reward function, all trained on data collected from interactions with the SMAC environment. A multi-agent policy then interacts with the simulator to obtain trajectories. These trajectories are then used to optimise the policy, but with a behaviour regularisation term towards the simulated trajectories. Evaluations on the SMAC benchmark show improved win rates and returns over existing online and offline multi-agent RL baselines. The paper also ablates the various components of the proposed method and provides visualisation of the simulated trajectories and their close match with actual trajectories.

Strengths

Proposes a method of training multi-agent RL using a learned simulator. It carefully considers issues such as data collection (by creating a method of generating images from stored game state), image tokenization, using inverse RL to learn the reward function, and regularising the policy update. As a result they achieve quite impressive performance gains in SMAC when compared with prior work. The paper thoroughly ablates the different components of their method and shows that while somewhat complex, each component is necessary for the best performance.

Weaknesses

The paper only studies SMAC and no other multi-agent domain. It is hard to evaluate the generality of the method without other evaluations on standard benchmarks. Adding evaluation results from one or two more domains would help strengthen the generality of this method. The write-up is a bit difficult to follow and contains a number of confusing parts (see also questions). The main figure, Figure 3, is never discussed in the main text. The part of Figure 1 on GPT-4 is not necessary for the paper. A significant amount of detail is located in the Appendix, for example Appendix B on how the training data for the simulation components are collected, and Appendix E on the experimental details. Consider moving these sections into the main text. The paper does not explain the methods used to collect the training data (EMC and IIE). To be better self contained, please give short explanations of these methods in the main text. Are they particularly tuned to the SMAC environment? What would have to change in domains where an effective exploration strategy is unknown?

Questions

In the input notation of the dynamics model why is L_t time-varying across a trajectory? From the text L is constant across an entire trajectory, could the input contain more than 1 trajectory? The notation could be improved. Could you explain in more detail how hindsight-relabelling is done and why it is necessary for reward function training? Is it the same or different from the text label used on the same set of collected trajectories as in dynamics model training? Is it only for data collected from the simulator? Clarifications for equation (1): What is D? Why does the f term have \tau as an input and an index i? Where do they come from? Is there a missing sum(s)? In the description it says that term is “over all possible actions”, what does this mean? Its inputs are trajectories and aren’t guaranteed to contain all actions? Why did the authors choose to only focus on a single multi-agent domain and not evaluate the method on any other, even toy, multi-agent benchmarks? Why do some tables report win rate and some rewards? Figure 8, does LBI timesteps account for the number of time steps in the data collected to train the simulator components? Would it be more accurate to shift the LBI curve to the right by that amount?

Rating

5

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

Limitations have been adequately addressed.

Reviewer sAvh5/10 · confidence 4/52024-07-13

Summary

This paper introduces a pipeline called Learning before Interaction, a approach to improve generative models' answers for multi-agent decision-making problems. The key idea is to integrate a language-guided simulator into the multi-agent reinforcement learning pipeline. The main contributions are - Creating VisionSMAC, a dataset that pairs states with images and task descriptions for the StarCraft Multi-Agent Challenge - Training an interactive simulator consisting of a dynamics model and a reward model The pipeline use the simulator to train a multi-agent policy through simulated interactions and generate answers by running the trained policy on the simulator.

Strengths

- LBI outperforms existing offline and imitation learning methods on unseen tasks in the SMAC. (good zero-shot generalization ability) - The approach produces explainable reward functions at critical states, which can help in understanding the decision-making process

Weaknesses

- (Not mentioned in the paper) Training an interactive simulator with both dynamics and reward models likely requires significant computational resources - It would be great if there are environments other than SMAC tested. Otherwise, the scope might be too narrow. - The approach requires a substantial amount of offline data to train the world model and compared with online MARL methods which is not fair comparison. The offline data is also not always available. - The ablation study shows that using image reference doesn't significantly improve performance, suggesting that this aspect of the model might not be fully utilized or optimized.

Questions

- What is the world model architecture and how do you train it? - Could you please write an algorithmic block for both inference and training? I found there are still many processes unclear after reading the paper. It will be very helpful if you can upload a sample code (please be careful about the anonymity in the code).

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

The authors do not mention a limitation.

Reviewer Frvn6/10 · confidence 4/52024-07-15

Summary

The paper presents a novel approach called Learning before Interaction (LBI) that integrates a language-guided simulator into the multi-agent reinforcement learning process to enhance the quality of solutions for complex decision-making problems. The LBI paradigm uses a world model comprising a dynamics model and a reward model to simulate interactions and improve policy learning. The authors introduce new offline MARL datasets for the StarCraft Multi-Agent Challenge and demonstrate LBI's effectiveness through notable performance on training and unseen tasks.

Strengths

1. Endeavors in building a feasible task-guided multi-agent world model. By introducing task description containing thorough information and the image of states of the task, this work empirically manage to build a world model being adaptive to various tasks in the training data distribution. 2. Designs for separate reward model and dynamics model. By decoupling the prediction of future observations and the prediction of rewards, this work can learn an efficient reward function involving reasonable credit assignments by adopting the inverse RL.

Weaknesses

1. Trivial framework designed for multi-agent world model. Overall, the method proposed in this work seems not to capture the distinct characteristics in the context of multi-agent scenarios as compared to single-agent scenarios. The built world model appears to be based on reduce multi-agent problems into a single-agent problem with joint observations and joint actions. What are the benefits of this approach? Why is it necessary to adopt this simplification? The article does not provide specific evidence or thorough analysis to support these architecture design decisions. 2. Lack of the investigation and citation of related multi-agent world model literature. Given that the method proposed in this work derives the optimal policy via learning in the imaginations of the world model, it is essential to elucidate the differences between this approach and those presented in [1, 2, 3, 4] and to conduct performance comparisons. Notably, the architecture closely resembles that of [1] in the absence of reward prediction. However, the entire paper lacks citations and clarifications regarding these similarities. 3. Unfair comparisons. According to the contents in the appendix, the average returns are almost 20 in all offline dataset of ten maps, which are nearly optimal in SMAC environments since the maximum return of an episode is 20 in almost all SMAC maps. Thus with a pretrained world model which has full access to the optimal demonstrations, the comparison with online methods is definitely not meaningful and convincing at all. 4. Heavily handcrafted processing. “Since it is possible for specific agents to become inactive before the game terminates, we mark the terminated timestep for each agent and enemy once its predicted health is less than zero …” To identify the predicted health, the algorithm has implicitly already known the specific meaning of the value in some certain dimensions in the observation, which is obviously induced from the prior knowledge or predefined rules from human beings. Unfortunately, such an operation makes this work only narrowed in the domain of SMAC. [1] Micheli, Vincent, Eloi Alonso, and François Fleuret. "Transformers are Sample-Efficient World Models." The Eleventh International Conference on Learning Representations. [2] Egorov, Vladimir, and Alexei Shpilman. "Scalable Multi-Agent Model-Based Reinforcement Learning." Proceedings of the 21st International Conference on Autonomous Agents and Multiagent Systems. 2022. [3] Zhang, Yang, et al. "Decentralized Transformers with Centralized Aggregation are Sample-Efficient Multi-Agent World Models." arXiv preprint arXiv:2406.15836 (2024). [4] Liu, Qihan, et al. "Efficient Multi-agent Reinforcement Learning by Planning." The Twelfth International Conference on Learning Representations.

Questions

See weaknesses.

Rating

6

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

N/A

Reviewer Frvn2024-08-10

Response

1. First, different from the single-agent scenario, the scalability issue stemming from increasing agents and the non-stationarity issue arising from the continuously changing environment of individual agents, are inherent in the multi-agent dynamics. However, the authors gave little explanation on whether and how this work can tackle these challenges. 2. The author may not carefully investigate the related works [1-2]. MAMBA and MARIE adopted a learning policy in the imagination scheme, enabling them to seamlessly incorporate different MARL methods. Their induced policies also remained decentralized. 3. I don't think the current experiment is aligned to demonstrate that training a policy in the world model is faster than training it in a real environment. In contrast, like how IRIS [3] delivered this similar insight via its experiments, it would be better to carry out experiments with a low data regime like Atari-100K to demonstrate the aim. 4. In terms of the handcrafted processing for obtaining agent death masks, it makes the work only applicable to the limited SMAC domain, weakening the contribution. Moreover, a prediction towards the agent death masks in the world model can better validate how accurate the world model, as the world model needs to understand the underlying game mechanics in SMAC. The handcrafted processing oversimplifies the learning. [1] Egorov, Vladimir, and Alexei Shpilman. Scalable Multi-Agent Model-Based Reinforcement Learning. [2] Zhana, Yang, et al. Decentralized Transformers with centralized Aggregation are Sample-Efficient Multi-Agent World Models [3] Micheli. Vincent, Eloi Alonso, and Francois Fleuret, Transformers are Sample-Efficient World Models

Authorsrebuttal2024-08-12

Rebuttal during the discussion

We appreciate the reviewer’s feedback! **1. The non-stationarity and the scalability issue.** We provide the state and joint actions as the input to train the world model (avoid the non-stationarity) and utilize the Transformer model (GPT2-like dynamics and Bert-like reward models) to provide enough representational capacity, where their attention mechanism performs feature selection and aggregation (for the scalability). To further elucidate our approach, we list the following paradigms: *(1) Decentralized training with decentralized execution*: While this paradigm excels in scalability during execution, it often leads to non-stationarity as agents treat others as part of the environment. *(2) Centralized training with decentralized execution*: It maintains execution scalability while introducing a centralized mechanism (e.g., a mixing network to aggregate the individual Q-values) to mitigate the non-stationarity problem. *(3) Communication-enabled training and execution*: Each agent learns and executes its policy based on their observations and (weighted) messages from others (e.g., actions, observations). This paradigm creates a trade-off between interaction complexity and representational capacity, a great compromise for non-stationarity and scalability issues. *(4) Centralized training with centralized execution*: While this paradigm eliminates the non-stationarity issue, it suffers from limited scalability during execution. Our world model serves as the environment, which leverages the Transformer model’s attention mechanism to capture interdependencies among agents and timesteps, similar to communication-enabled paradigms (e.g., MAT [1]). The concurrent work [2] also proposes a Perceiver Transformer as a communication module, which utilizes the querying and self-attention mechanisms to perform agent-wise aggregation, transforming the joint representation sequence (joint observations and actions) into lower-dimensional features. The difference is that our world model integrates these processes into the large transformer model. **2&3. MAMBA and MARIE, and the low data regime.** We acknowledge that MAMBA [3] and MARIE [2] are indeed Dyna-style algorithms that learn policy in the imagination, and their induced policies are decentralized. We would like to clarify our stance: * MAMBA and MARIE assume that the reward function is presented in the replay buffer. They train the reward predictor using log-likelihood and cross-entropy loss, respectively. **Conversely**, we suppose the reward is unknown in the offline dataset. We construct the reward function using a bidirectional Transformer, which maximizes the likelihood of trajectories in hindsight relabeled expert demonstrations. * MAMBA and MARIE are online model-based RL methods. They explore the environment using a policy, store the collected data in a replay buffer, update the world model with samples from this buffer, and use rollouts from the world model to refine the policy. They can benefit from interacting with the real environment, allowing them to update and rectify their world models continuously - e.g., if agents imagine trajectories in out-of-distribution regions that lead to suboptimal updates, they can correct this during exploration by gathering more data in these regions. Consequently, they conducted experiments in low-data regimes like Atari-100K to show their data efficiency. **In contrast**, our dynamics model is trained in an offline manner via a static, offline dataset. Unlike online methods, no exploration is available for the model to rectify itself, leading to the distributional shift issue - where the deployment scenario may differ significantly from the historical dataset regarding state and action space. To mitigate this, we employ a behavior regularization term to encourage exploitation, specifically by discouraging agents from executing out-of-distribution behaviors during inference. **4. The agent death masks.** Thank you for your insightful suggestion. We would like to use a binary gate function for each agent to replace the handcrafted death masks and learn it via binary cross-entropy loss in a supervised manner. We acknowledge the potential for further generalization. However, we respectfully disagree with the notion that the current approach oversimplifies the learning process. As shown in Fig.4, the learned world model captures the dynamics and learns a meaningful reward function even in this complex scenario. This demonstrates that the model correctly understands the transitions within the game, validating its ability to grasp the underlying mechanics of SMAC. [1] Wen, Muning, et al. Multi-agent reinforcement learning is a sequence modeling problem. [2] Zhana, Yang, et al. Decentralized Transformers with centralized Aggregation are Sample-Efficient Multi-Agent World Models. [3] Micheli. Vincent, Eloi Alonso, and Francois Fleuret, Transformers are Sample-Efficient World Models.

Reviewer sAvh2024-08-11

Thanks for the comments. I am happy to raise the score.

Reviewer 2fu72024-08-13

Reply

I thank the authors for the detailed rebuttal. Most of my concerns has been addressed. However, one key concern of the limiting scope of the work only on the SMAC benchmark (which it seems is also shared with other reviewers) has not been addressed. I will maintain my original rating.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC