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.

Many real-world problems, such as controlling swarms of drones and urban traffic, naturally lend themselves to modeling as multi-agent reinforcement learning (RL) problems. However, existing multi-agent RL methods often suffer from scalability challenges, primarily due to the introduction of communication among agents. Consequently, a key challenge lies in adapting the success of deep learning in single-agent RL to the multi-agent setting. In response to this challenge, we propose an approach that fundamentally reimagines multi-agent environments. Unlike conventional methods that model each agent individually with separate networks, our approach, the Bottom Up Network (BUN), adopts a unique perspective. BUN treats the collective of multi-agents as a unified entity while employing a specialized weight initialization strategy that promotes independent learning. Furthermore, we dynamically establish connections among agents using gradient information, enabling coordination when necessary while maintaining these connections as limited and sparse to effectively manage the computational budget. Our extensive empirical evaluations across a variety of cooperative multi-agent scenarios, including tasks such as cooperative navigation and traffic control, consistently demonstrate BUN's superiority over baseline methods with substantially reduced computational costs.

With expansive state-action spaces, efficient multi-agent exploration remains a longstanding challenge in reinforcement learning. Although pursuing novelty, diversity, or uncertainty attracts increasing attention, redundant efforts brought by exploration without proper guidance choices poses a practical issue for the community. This paper introduces a systematic approach, termed LEMAE, choosing to channel informative task-relevant guidance from a knowledgeable Large Language Model (LLM) for Efficient Multi-Agent Exploration. Specifically, we ground linguistic knowledge from LLM into symbolic key states, that are critical for task fulfillment, in a discriminative manner at low LLM inference costs. To unleash the power of key states, we design Subspace-based Hindsight Intrinsic Reward (SHIR) to guide agents toward key states by increasing reward density. Additionally, we build the Key State Memory Tree (KSMT) to track transitions between key states in a specific task for organized exploration. Benefiting from diminishing redundant explorations, LEMAE outperforms existing SOTA approaches on the challenging benchmarks (e.g., SMAC and MPE) by a large margin, achieving a 10x acceleration in certain scenarios.

Recent advancements in large language model (LLM)-powered agents have shown that collective intelligence can significantly outperform individual capabilities, largely attributed to the meticulously designed inter-agent communication topologies. Though impressive in performance, existing multi-agent pipelines inherently introduce substantial token overhead, as well as increased economic costs, which pose challenges for their large-scale deployments. In response to this challenge, we propose an economical, simple, and robust multi-agent communication framework, termed $\texttt{AgentPrune}$, which can seamlessly integrate into mainstream multi-agent systems and prunes redundant or even malicious communication messages. Technically, $\texttt{AgentPrune}$ is the first to identify and formally define the \textit{communication redundancy} issue present in current LLM-based multi-agent pipelines, and efficiently performs one-shot pruning on the spatial-temporal message-passing graph, yielding a token-economic and high-performing communication topology. Extensive experiments across six benchmarks demonstrate that $\texttt{AgentPrune}$ \textbf{(I)} achieves comparable results as state-of-the-art topologies at merely $\$5.6$ cost compared to their $\$43.7$, \textbf{(II)} integrates seamlessly into existing multi-agent frameworks with $28.1\%\sim72.8\%\downarrow$ token reduction, and \textbf{(III)} successfully defend against two types of agent-based adversarial attacks with $3.5\%\sim10.8\%\uparrow$ performance boost.

Offline reinforcement learning (RL) has garnered significant attention for its ability to learn effective policies from pre-collected datasets without the need for further environmental interactions. While promising results have been demonstrated in single-agent settings, offline multi-agent reinforcement learning (MARL) presents additional challenges due to the large joint state-action space and the complexity of multi-agent behaviors. A key issue in offline RL is the distributional shift, which arises when the target policy being optimized deviates from the behavior policy that generated the data. This problem is exacerbated in MARL due to the interdependence between agents' local policies and the expansive joint state-action space. Prior approaches have primarily addressed this challenge by incorporating regularization in the space of either Q-functions or policies. In this work, we introduce a regularizer in the space of stationary distributions to better handle distributional shift. Our algorithm, ComaDICE, offers a principled framework for offline cooperative MARL by incorporating stationary distribution regularization for the global learning policy, complemented by a carefully structured multi-agent value decomposition strategy to facilitate multi-agent training. Through extensive experiments on the multi-agent MuJoCo and StarCraft II benchmarks, we demonstrate that ComaDICE achieves superior performance compared to state-of-the-art offline MARL methods across nearly all tasks. Over the years, deep RL has achieved remarkable success in various decision-making tasks (Levine et al., 2016; Silver et al., 2017; Kalashnikov et al., 2018; Haydari & Yılmaz, 2020). However, a significant limitation of deep RL is its need for millions of interactions with the environment to gather experiences for policy improvement.

Reconstructing past events requires reasoning across long time horizons. To figure out what happened, we need to use our prior knowledge about the world and human behavior and draw inferences from various sources of evidence including visual, language, and auditory cues. We introduce MARPLE, a benchmark for evaluating long-horizon inference capabilities using multi-modal evidence. Our benchmark features agents interacting with simulated households, supporting vision, language, and auditory stimuli, as well as procedurally generated environments and agent behaviors. Inspired by classic ``whodunit'' stories, we ask AI models and human participants to infer which agent caused a change in the environment based on a step-by-step replay of what actually happened. The goal is to correctly identify the culprit as early as possible. Our findings show that human participants outperform both traditional Monte Carlo simulation methods and an LLM baseline (GPT-4) on this task. Compared to humans, traditional inference models are less robust and performant, while GPT-4 has difficulty comprehending environmental changes. We analyze what factors influence inference performance and ablate different modes of evidence, finding that all modes are valuable for performance. Overall, our experiments demonstrate that the long-horizon, multimodal inference tasks in our benchmark present a challenge to current models.

Traditional multi-agent path finding (MAPF) methods try to compute entire start-goal paths which are collision free. However, computing an entire path can take too long for MAPF systems where agents need to replan fast. Methods that address this typically employ a "windowed" approach and only try to find collision free paths for a small windowed timestep horizon. This adaptation comes at the cost of incompleteness; all current windowed approaches can become stuck in deadlock or livelock. Our main contribution is to introduce our framework, WinC-MAPF, for Windowed MAPF that enables completeness. Our framework uses heuristic update insights from single-agent real-time heuristic search algorithms as well as agent independence ideas from MAPF algorithms. We also develop Single-Step CBS (SS-CBS), an instantiation of this framework using a novel modification to CBS. We show how SS-CBS, which only plans a single step and updates heuristics, can effectively solve tough scenarios where existing windowed approaches fail.

As the field of multi-agent reinforcement learning (MARL) progresses towards larger and more complex environments, achieving strong performance while maintaining memory efficiency and scalability to many agents becomes increasingly important. Although recent research has led to several advanced algorithms, to date, none fully address all of these key properties simultaneously. In this work, we introduce Sable, a novel and theoretically sound algorithm that adapts the retention mechanism from Retentive Networks to MARL. Sable's retention-based sequence modelling architecture allows for computationally efficient scaling to a large number of agents, as well as maintaining a long temporal context, making it well-suited for large-scale partially observable environments. Through extensive evaluations across six diverse environments, we demonstrate how Sable is able to significantly outperform existing state-of-the-art methods in the majority of tasks (34 out of 45, roughly 75%). Furthermore, Sable demonstrates stable performance as we scale the number of agents, handling environments with more than a thousand agents while exhibiting a linear increase in memory usage. Finally, we conduct ablation studies to isolate the source of Sable's performance gains and confirm its efficient computational memory usage. Our results highlight Sable's performance and efficiency, positioning it as a leading approach to MARL at scale. When considering large-scale practical applications of multi-agent reinforcement learning (MARL) such as autonomous driving (Lian & Deshmukh, 2006; Zhou et al., 2021; Li et al., 2022) and electricity grid control (Kamboj et al., 2011; Li et al., 2016), it becomes increasingly important to maintain three key properties for a system to be effective: strong performance, memory efficiency, and scalability to many agents. Although many existing MARL approaches exhibit one or two of these properties, a solution effectively encompassing all three remains elusive. To briefly illustrate our point, we consider the spectrum of approaches to MARL. Such algorithms demonstrate proficiency in handling many agents in a memory efficient way by typically using shared parameters and conditioning on an agent identifier. However, at scale, the performance of fully decentralised methods remains suboptimal compared to more centralised approaches (Papoudakis et al., 2021; Yu et al., 2022; Wen et al., 2022). Between decentralised and centralised methods, lie CTDE approaches (Lowe et al., 2017; Papoudakis et al., 2021; Yu et al., 2022).

Effective environment perception is crucial for enabling downstream robotic applications. Individual robotic agents often face occlusion and limited visibility issues, whereas multi-agent systems can offer a more comprehensive mapping of the environment, quicker coverage, and increased fault tolerance. In this paper, we propose a collaborative multi-agent perception system where agents collectively learn a neural radiance field (NeRF) from posed RGB images to represent a scene. Each agent processes its local sensory data and shares only its learned NeRF model with other agents, reducing communication overhead. Given NeRF's low memory footprint, this approach is well-suited for robotic systems with limited bandwidth, where transmitting all raw data is impractical. Our distributed learning framework ensures consistency across agents' local NeRF models, enabling convergence to a unified scene representation. We show the effectiveness of our method through an extensive set of experiments on datasets containing challenging real-world scenes, achieving performance comparable to centralized mapping of the environment where data is sent to a central server for processing. Additionally, we find that multi-agent learning provides regularization benefits, improving geometric consistency in scenarios with sparse input views. We show that in such scenarios, multi-agent mapping can even outperform centralized training.

Many decision-making problems feature multiple objectives where it is not always possible to know the preferences of a human or agent decision-maker for different objectives. However, demonstrated behaviors from the decision-maker are often available. This research proposes a dynamic weight-based preference inference (DWPI) algorithm that can infer the preferences of agents acting in multi-objective decision-making problems from demonstrations. The proposed algorithm is evaluated on three multi-objective Markov decision processes: Deep Sea Treasure, Traffic, and Item Gathering, and is compared to two existing preference inference algorithms. Empirical results demonstrate significant improvements compared to the baseline algorithms, in terms of both time efficiency and inference accuracy. The DWPI algorithm maintains its performance when inferring preferences for sub-optimal demonstrations. Moreover, the DWPI algorithm does not necessitate any interactions with the user during inference - only demonstrations are required. We provide a correctness proof and complexity analysis of the algorithm and statistically evaluate the performance under different representation of demonstrations.