Training multiple agents to coordinate is an important problem with applications in robotics, game theory, economics, and social sciences. However, most existing Multi-Agent Reinforcement Learning (MARL) methods are online and thus impractical for real-world applications in which collecting new interactions is costly or dangerous. While these algorithms should leverage offline data when available, doing so gives rise to the offline coordination problem. Specifically, we identify and formalize the strategy agreement (SA) and the strategy fine-tuning (SFT) challenges, two coordination issues at which current offline MARL algorithms fail. To address this setback, we propose a simple model-based approach that generates synthetic interaction data and enables agents to converge on a strategy while fine-tuning their policies accordingly. Our resulting method, Model-based Offline Multi-Agent Proximal Policy Optimization (MOMA-PPO), outperforms the prevalent learning methods in challenging offline multi-agent MuJoCo tasks even under severe partial observability and with learned world models.

Centralized Training with Decentralized Execution (CTDE) has recently emerged as a popular framework for cooperative Multi-Agent Reinforcement Learning (MARL), where agents can use additional global state information to guide training in a centralized way and make their own decisions only based on decentralized local policies. Despite the encouraging results achieved, CTDE makes an independence assumption on agent policies, which limits agents from adopting global cooperative information from each other during centralized training. Therefore, we argue that the existing CTDE framework cannot fully utilize global information for training, leading to an inefficient joint-policy exploration and even suboptimal results. In this paper, we introduce a novel Centralized Advising and Decentralized Pruning (CADP) framework for multi-agent reinforcement learning, that not only enables an efficacious message exchange among agents during training but also guarantees the independent policies for execution. Firstly, CADP endows agents the explicit communication channel to seek and take advice from different agents for more centralized training. To further ensure the decentralized execution, we propose a smooth model pruning mechanism to progressively constrain the agent communication into a closed one without degradation in agent cooperation capability. Empirical evaluations on StarCraft II micromanagement challenge and Google Research Football benchmarks and and across different MARL backbones demonstrate that the proposed framework achieves superior performance compared with the state-of-the-art counterparts. Our code is available at https://github.com/zyh1999/CADP.

Most existing works consider direct perturbations of victim's state/action or the underlying transition dynamics to show vulnerability of reinforcement learning agents under adversarial attacks. However, such direct manipulation may not always be feasible in practice. In this paper, we consider another common and realistic attack setup: in a multi-agent RL setting with well-trained agents, during deployment time, the victim agent $\nu$ is exploited by an attacker who controls another agent $\alpha$ to act adversarially against the victim using an \textit{adversarial policy}. Prior attack models under such setup do not consider that the attacker can confront resistance and thus can only take partial control of the agent $\alpha$, as well as introducing perceivable ``abnormal'' behaviors that are easily detectable. A provable defense against these adversarial policies is also lacking. To resolve these issues, we introduce a more general attack formulation that models to what extent the adversary is able to control the agent to produce the adversarial policy. Based on such a generalized attack framework, the attacker can also regulate the state distribution shift caused by the attack through an attack budget, and thus produce stealthy adversarial policies that can exploit the victim agent. Furthermore, we provide the first provably robust defenses with convergence guarantee to the most robust victim policy via adversarial training with timescale separation, in sharp contrast to adversarial training in supervised learning which may only provide {\it empirical} defenses.

Mean Field Control (MFC) is a powerful approximation tool to solve large-scale Multi-Agent Reinforcement Learning (MARL) problems. However, the success of MFC relies on the presumption that given the local states and actions of all the agents, the next (local) states of the agents evolve conditionally independent of each other. Here we demonstrate that even in a MARL setting where agents share a common global state in addition to their local states evolving conditionally independently (thus introducing a correlation between the state transition processes of individual agents), the MFC can still be applied as a good approximation tool. The global state is assumed to be non-decomposable i.e., it cannot be expressed as a collection of local states of the agents. We compute the approximation error as $\mathcal{O}(e)$ where $e=\frac{1}{\sqrt{N}}\left[\sqrt{|\mathcal{X}|} +\sqrt{|\mathcal{U}|}\right]$. The size of the agent population is denoted by the term $N$, and $|\mathcal{X}|, |\mathcal{U}|$ respectively indicate the sizes of (local) state and action spaces of individual agents. The approximation error is found to be independent of the size of the shared global state space. We further demonstrate that in a special case if the reward and state transition functions are independent of the action distribution of the population, then the error can be improved to $e=\frac{\sqrt{|\mathcal{X}|}}{\sqrt{N}}$. Finally, we devise a Natural Policy Gradient based algorithm that solves the MFC problem with $\mathcal{O}(\epsilon^{-3})$ sample complexity and obtains a policy that is within $\mathcal{O}(\max\{e,\epsilon\})$ error of the optimal MARL policy for any $\epsilon>0$.

Autonomous systems often have logical constraints arising, for example, from safety, operational, or regulatory requirements. Such constraints can be expressed using temporal logic specifications. The system state is often partially observable. Moreover, it could encompass a team of multiple agents with a common objective but disparate information structures and constraints. In this paper, we first introduce an optimal control theory for partially observable Markov decision processes (POMDPs) with finite linear temporal logic constraints. We provide a structured methodology for synthesizing policies that maximize a cumulative reward while ensuring that the probability of satisfying a temporal logic constraint is sufficiently high. Our approach comes with guarantees on approximate reward optimality and constraint satisfaction. We then build on this approach to design an optimal control framework for logically constrained multi-agent settings with information asymmetry. We illustrate the effectiveness of our approach by implementing it on several case studies.

We study the problem of multi-agent coordination in unpredictable and partially observable environments, that is, environments whose future evolution is unknown a priori and that can only be partially observed. We are motivated by the future of autonomy that involves multiple robots coordinating actions in dynamic, unstructured, and partially observable environments to complete complex tasks such as target tracking, environmental mapping, and area monitoring. Such tasks are often modeled as submodular maximization coordination problems due to the information overlap among the robots. We introduce the first submodular coordination algorithm with bandit feedback and bounded tracking regret -- bandit feedback is the robots' ability to compute in hindsight only the effect of their chosen actions, instead of all the alternative actions that they could have chosen instead, due to the partial observability; and tracking regret is the algorithm's suboptimality with respect to the optimal time-varying actions that fully know the future a priori. The bound gracefully degrades with the environments' capacity to change adversarially, quantifying how often the robots should re-select actions to learn to coordinate as if they fully knew the future a priori. The algorithm generalizes the seminal Sequential Greedy algorithm by Fisher et al. to the bandit setting, by leveraging submodularity and algorithms for the problem of tracking the best action. We validate our algorithm in simulated scenarios of multi-target tracking.

Prototyping and validating hardware-software components, sub-systems and systems within the intelligent transportation system-of-systems framework requires a modular yet flexible and open-access ecosystem. This work presents our attempt towards developing such a comprehensive research and education ecosystem, called AutoDRIVE, for synergistically prototyping, simulating and deploying cyber-physical solutions pertaining to autonomous driving as well as smart city management. AutoDRIVE features both software as well as hardware-in-the-loop testing interfaces with openly accessible scaled vehicle and infrastructure components. The ecosystem is compatible with a variety of development frameworks, and supports both single and multi-agent paradigms through local as well as distributed computing. Most critically, AutoDRIVE is intended to be modularly expandable to explore emergent technologies, and this work highlights various complementary features and capabilities of the proposed ecosystem by demonstrating four such deployment use-cases: (i) autonomous parking using probabilistic robotics approach for mapping, localization, path planning and control; (ii) behavioral cloning using computer vision and deep imitation learning; (iii) intersection traversal using vehicle-to-vehicle communication and deep reinforcement learning; and (iv) smart city management using vehicle-to-infrastructure communication and internet-of-things.

Software bots operating in multiple virtual digital platforms must understand the platforms' affordances and behave like human users. Platform affordances or features differ from one application platform to another or through a life cycle, requiring such bots to be adaptable. Moreover, bots in such platforms could cooperate with humans or other software agents for work or to learn specific behavior patterns. However, present-day bots, particularly chatbots, other than language processing and prediction, are far from reaching a human user's behavior level within complex business information systems. They lack the cognitive capabilities to sense and act in such virtual environments, rendering their development a challenge to artificial general intelligence research. In this study, we problematize and investigate assumptions in conceptualizing software bot architecture by directing attention to significant architectural research challenges in developing cognitive bots endowed with complex behavior for operation on information systems. As an outlook, we propose alternate architectural assumptions to consider in future bot design and bot development frameworks.

The reliability of current autonomous driving systems is often jeopardized in situations when the vehicle's field-of-view is limited by nearby occluding objects. To mitigate this problem, vehicle-to-vehicle communication to share sensor information among multiple autonomous driving vehicles has been proposed. However, to enable timely processing and use of shared sensor data, it is necessary to constrain communication bandwidth, and prior work has done so by restricting the number of other cooperative vehicles and randomly selecting the subset of vehicles to exchange information with from all those that are within communication range. Although simple and cost effective from a communication perspective, this selection approach suffers from its susceptibility to missing those vehicles that possess the perception information most critical to navigation planning. Inspired by recent multi-agent path finding research, we propose a novel selective communication algorithm for cooperative perception to address this shortcoming. Implemented with a lightweight perception network and a previously developed control network, our algorithm is shown to produce higher success rates than a random selection approach on previously studied safety-critical driving scenario simulations, with minimal additional communication overhead.

Great storytellers know how to take us on a journey. They direct characters to act -- not necessarily in the most rational way -- but rather in a way that leads to interesting situations, and ultimately creates an impactful experience for audience members looking on. If audience experience is what matters most, then can we help artists and animators *directly* craft such experiences, independent of the concrete character actions needed to evoke those experiences? In this paper, we offer a novel computational framework for such tools. Our key idea is to optimize animations with respect to *simulated* audience members' experiences. To simulate the audience, we borrow an established principle from cognitive science: that human social intuition can be modeled as "inverse planning," the task of inferring an agent's (hidden) goals from its (observed) actions. Building on this model, we treat storytelling as "*inverse* inverse planning," the task of choosing actions to manipulate an inverse planner's inferences. Our framework is grounded in literary theory, naturally capturing many storytelling elements from first principles. We give a series of examples to demonstrate this, with supporting evidence from human subject studies.