Intrinsic rewards have been increasingly used to mitigate the sparse reward problem in single-agent reinforcement learning. These intrinsic rewards encourage the agent to look for novel experiences, guiding the agent to explore the environment sufficiently despite the lack of extrinsic rewards. Curiosity-driven exploration is a simple yet efficient approach that quantifies this novelty as the prediction error of the agent's curiosity module, an internal neural network that is trained to predict the agent's next state given its current state and action. We show here, however, that naively using this curiosity-driven approach to guide exploration in sparse reward cooperative multi-agent environments does not consistently lead to improved results. Straightforward multi-agent extensions of curiosity-driven exploration take into consideration either individual or collective novelty only and thus, they do not provide a distinct but collaborative intrinsic reward signal that is essential for learning in cooperative multi-agent tasks. In this work, we propose a curiosity-driven multi-agent exploration method that has the mixed objective of motivating the agents to explore the environment in ways that are individually and collectively novel. First, we develop a two-headed curiosity module that is trained to predict the corresponding agent's next observation in the first head and the next joint observation in the second head. Second, we design the intrinsic reward formula to be the sum of the individual and joint prediction errors of this curiosity module. We empirically show that the combination of our curiosity module architecture and intrinsic reward formulation guides multi-agent exploration more efficiently than baseline approaches, thereby providing the best performance boost to MARL algorithms in cooperative navigation environments with sparse rewards.

Large sequence model (SM) such as GPT series and BERT has displayed outstanding performance and generalization capabilities on vision, language, and recently reinforcement learning tasks. A natural follow-up question is how to abstract multi-agent decision making into an SM problem and benefit from the prosperous development of SMs. In this paper, we introduce a novel architecture named Multi-Agent Transformer (MAT) that effectively casts cooperative multi-agent reinforcement learning (MARL) into SM problems wherein the task is to map agents' observation sequence to agents' optimal action sequence. Our goal is to build the bridge between MARL and SMs so that the modeling power of modern sequence models can be unleashed for MARL. Central to our MAT is an encoder-decoder architecture which leverages the multi-agent advantage decomposition theorem to transform the joint policy search problem into a sequential decision making process; this renders only linear time complexity for multi-agent problems and, most importantly, endows MAT with monotonic performance improvement guarantee. Unlike prior arts such as Decision Transformer fit only pre-collected offline data, MAT is trained by online trials and errors from the environment in an on-policy fashion. To validate MAT, we conduct extensive experiments on StarCraftII, Multi-Agent MuJoCo, Dexterous Hands Manipulation, and Google Research Football benchmarks. Results demonstrate that MAT achieves superior performance and data efficiency compared to strong baselines including MAPPO and HAPPO. Furthermore, we demonstrate that MAT is an excellent few-short learner on unseen tasks regardless of changes in the number of agents. See our project page at https://sites.google.com/view/multi-agent-transformer.

We present a game benchmark for testing human-swarm control algorithms and interfaces in a real-time, high-cadence scenario. Our benchmark consists of a swarm vs. swarm game in a virtual ROS environment in which the goal of the game is to capture all agents from the opposing swarm; the game's high-cadence is a result of the capture rules, which cause agent team sizes to fluctuate rapidly. These rules require players to consider both the number of agents currently at their disposal and the behavior of their opponent's swarm when they plan actions. We demonstrate our game benchmark with a default human-swarm control system that enables a player to interact with their swarm through a high-level touchscreen interface. The touchscreen interface transforms player gestures into swarm control commands via a low-level decentralized ergodic control framework. We compare our default human-swarm control system to a flocking-based control system, and discuss traits that are crucial for swarm control algorithms and interfaces operating in real-time, high-cadence scenarios like our game benchmark. Our game benchmark code is available on Github; more information can be found at https://sites.google.com/view/swarm-game-benchmark.

Value decomposition multi-agent reinforcement learning methods learn the global value function as a mixing of each agent's individual utility functions. Coordination graphs (CGs) represent a higher-order decomposition by incorporating pairwise payoff functions and thus is supposed to have a more powerful representational capacity. However, CGs decompose the global value function linearly over local value functions, severely limiting the complexity of the value function class that can be represented. In this paper, we propose the first non-linear coordination graph by extending CG value decomposition beyond the linear case. One major challenge is to conduct greedy action selections in this new function class to which commonly adopted DCOP algorithms are no longer applicable. We study how to solve this problem when mixing networks with LeakyReLU activation are used. An enumeration method with a global optimality guarantee is proposed and motivates an efficient iterative optimization method with a local optimality guarantee. We find that our method can achieve superior performance on challenging multi-agent coordination tasks like MACO.

We investigate the use of natural language to drive the generalization of policies in multi-agent settings. Unlike single-agent settings, the generalization of policies should also consider the influence of other agents. Besides, with the increasing number of entities in multi-agent settings, more agent-entity interactions are needed for language grounding, and the enormous search space could impede the learning process. Moreover, given a simple general instruction, e.g., beating all enemies, agents are required to decompose it into multiple subgoals and figure out the right one to focus on. Inspired by previous work, we try to address these issues at the entity level and propose a novel framework for language grounding in multi-agent reinforcement learning, entity divider (EnDi). EnDi enables agents to independently learn subgoal division at the entity level and act in the environment based on the associated entities. The subgoal division is regularized by opponent modeling to avoid subgoal conflicts and promote coordinated strategies. Empirically, EnDi demonstrates the strong generalization ability to unseen games with new dynamics and expresses the superiority over existing methods.

This paper aims to tackle the interactive behavior prediction task, and proposes a novel Conditional Goal-oriented Trajectory Prediction (CGTP) framework to jointly generate scene-compliant trajectories of two interacting agents. Our CGTP framework is an end to end and interpretable model, including three main stages: context encoding, goal interactive prediction and trajectory interactive prediction. First, a Goals-of-Interest Network (GoINet) is designed to extract the interactive features between agent-to-agent and agent-to-goals using a graph-based vectorized representation. Further, the Conditional Goal Prediction Network (CGPNet) focuses on goal interactive prediction via a combined form of marginal and conditional goal predictors. Finally, the Goaloriented Trajectory Forecasting Network (GTFNet) is proposed to implement trajectory interactive prediction via the conditional goal-oriented predictors, with the predicted future states of the other interacting agent taken as inputs. In addition, a new goal interactive loss is developed to better learn the joint probability distribution over goal candidates between two interacting agents. In the end, the proposed method is conducted on Argoverse motion forecasting dataset, In-house cut-in dataset, and Waymo open motion dataset. The comparative results demonstrate the superior performance of our proposed CGTP model than the mainstream prediction methods.

Contemporary swarm indicators are often used in isolation, focused on extracting information at the individual or collective levels. Consequently, these are seldom integrated to infer a top-level operating picture of the swarm, its members, and its overall collective dynamics. The primary contribution of this paper is to organise a suite of indicators about swarms into an ontologically-arranged collection of information markers to characterise the swarm from the perspective of an external observer\textemdash, a recognition agent. Our contribution shows the foundations for a new area of research that we tile swarm analytics, whose primary concern is with the design and organisation of collections of swarm markers to understand, detect, recognise, track, and learn a particular insight about a swarm system. We present our designed framework of information markers that offer a new avenue for swarm research, especially for heterogeneous and cognitive swarms that may require more advanced capabilities to detect agencies and categorise agent influences and responses.

Deep reinforcement learning in continuous domains focuses on learning control policies that map states to distributions over actions that ideally concentrate on the optimal choices in each step. In multi-agent navigation problems, the optimal actions depend heavily on the agents' density. Their interaction patterns grow exponentially with respect to such density, making it hard for learning-based methods to generalize. We propose to switch the learning objectives from predicting the optimal actions to predicting sets of admissible actions, which we call control admissibility models (CAMs), such that they can be easily composed and used for online inference for an arbitrary number of agents. We design CAMs using graph neural networks and develop training methods that optimize the CAMs in the standard model-free setting, with the additional benefit of eliminating the need for reward engineering typically required to balance collision avoidance and goal-reaching requirements. We evaluate the proposed approach in multi-agent navigation environments. We show that the CAM models can be trained in environments with only a few agents and be easily composed for deployment in dense environments with hundreds of agents, achieving better performance than state-of-the-art methods.

The cooperative Multi-A gent R einforcement Learning (MARL) with permutation invariant agents framework has achieved tremendous empirical successes in real-world applications. Unfortunately, the theoretical understanding of this MARL problem is lacking due to the curse of many agents and the limited exploration of the relational reasoning in existing works. In this paper, we verify that the transformer implements complex relational reasoning, and we propose and analyze model-free and model-based offline MARL algorithms with the transformer approximators. We prove that the suboptimality gaps of the model-free and model-based algorithms are independent of and logarithmic in the number of agents respectively, which mitigates the curse of many agents. These results are consequences of a novel generalization error bound of the transformer and a novel analysis of the Maximum Likelihood Estimate (MLE) of the system dynamics with the transformer. Our model-based algorithm is the first provably efficient MARL algorithm that explicitly exploits the permutation invariance of the agents. Our improved generalization bound may be of independent interest and is applicable to other regression problems related to the transformer beyond MARL.

Multi-agent reinforcement learning has drawn increasing attention in practice, e.g., robotics and automatic driving, as it can explore optimal policies using samples generated by interacting with the environment. However, high reward uncertainty still remains a problem when we want to train a satisfactory model, because obtaining high-quality reward feedback is usually expensive and even infeasible. To handle this issue, previous methods mainly focus on passive reward correction. At the same time, recent active reward estimation methods have proven to be a recipe for reducing the effect of reward uncertainty. In this paper, we propose a novel Distributional Reward Estimation framework for effective Multi-Agent Reinforcement Learning (DRE-MARL). Our main idea is to design the multi-action-branch reward estimation and policy-weighted reward aggregation for stabilized training. Specifically, we design the multi-action-branch reward estimation to model reward distributions on all action branches. Then we utilize reward aggregation to obtain stable updating signals during training. Our intuition is that consideration of all possible consequences of actions could be useful for learning policies. The superiority of the DRE-MARL is demonstrated using benchmark multi-agent scenarios, compared with the SOTA baselines in terms of both effectiveness and robustness.