Cooperative multi-agent multi-armed bandits (CMA2B) consider the collaborative efforts of multiple agents in a shared multi-armed bandit game. We study latent vulnerabilities exposed by this collaboration and consider adversarial attacks on a few agents with the goal of influencing the decisions of the rest. More specifically, we study adversarial attacks on CMA2B in both homogeneous settings, where agents operate with the same arm set, and heterogeneous settings, where agents have distinct arm sets. In the homogeneous setting, we propose attack strategies that, by targeting just one agent, convince all agents to select a particular target arm $T-o(T)$ times while incurring $o(T)$ attack costs in $T$ rounds. In the heterogeneous setting, we prove that a target arm attack requires linear attack costs and propose attack strategies that can force a maximum number of agents to suffer linear regrets while incurring sublinear costs and only manipulating the observations of a few target agents. Numerical experiments validate the effectiveness of our proposed attack strategies.

Recent works have proven that intricate cooperative behaviors can emerge in agents trained using meta reinforcement learning on open ended task distributions using self-play. While the results are impressive, we argue that self-play and other centralized training techniques do not accurately reflect how general collective exploration strategies emerge in the natural world: through decentralized training and over an open-ended distribution of tasks. In this work we therefore investigate the emergence of collective exploration strategies, where several agents meta-learn independent recurrent policies on an open ended distribution of tasks. To this end we introduce a novel environment with an open ended procedurally generated task space which dynamically combines multiple subtasks sampled from five diverse task types to form a vast distribution of task trees. We show that decentralized agents trained in our environment exhibit strong generalization abilities when confronted with novel objects at test time. Additionally, despite never being forced to cooperate during training the agents learn collective exploration strategies which allow them to solve novel tasks never encountered during training. We further find that the agents learned collective exploration strategies extend to an open ended task setting, allowing them to solve task trees of twice the depth compared to the ones seen during training. Our open source code as well as videos of the agents can be found on our companion website.

Responsible AI has risen to the forefront of the AI research community. As neural network-based learning algorithms continue to permeate real-world applications, the field of Responsible AI has played a large role in ensuring that such systems maintain a high-level of human-compatibility. Despite this progress, the state of the art in Responsible AI has ignored one crucial point: human problems are multi-agent problems. Predominant approaches largely consider the performance of a single AI system in isolation, but human problems are, by their very nature, multi-agent. From driving in traffic to negotiating economic policy, human problem-solving involves interaction and the interplay of the actions and motives of multiple individuals. This dissertation develops the study of responsible emergent multi-agent behavior, illustrating how researchers and practitioners can better understand and shape multi-agent learning with respect to three pillars of Responsible AI: interpretability, fairness, and robustness. First, I investigate multi-agent interpretability, presenting novel techniques for understanding emergent multi-agent behavior at multiple levels of granularity. With respect to low-level interpretability, I examine the extent to which implicit communication emerges as an aid to coordination in multi-agent populations. I introduce a novel curriculum-driven method for learning high-performing policies in difficult, sparse reward environments and show through a measure of position-based social influence that multi-agent teams that learn sophisticated coordination strategies exchange significantly more information through implicit signals than lesser-coordinated agents. Then, at a high-level, I study concept-based interpretability in the context of multi-agent learning. I propose a novel method for learning intrinsically interpretable, concept-based policies and show that it enables...

Open multi-agent systems (MAS) are composed of agents under different organisational control, and whose internal goals and mental states cannot be observed. In such systems, agents often have differing capabilities, and must rely on each other when pursuing their goals, making task delegation commonplace. This delegation occurs when one agent (the delegator) requests that another (the delegatee), execute a task. A fundamental problem faced by the delegator involves selecting the most appropriate delegatee to whom the task should be delegated, and a significant body of work centred around trust and reputation systems has examined how such a delegation decision should take place [6, 12, 13]. At their heart, trust and reputation systems associate a rating with each potential delegatee, and select who to delegate a task to based on this rating. Following task execution, the rating is updated based on how well the task was completed. Different systems compute the ratings differently, for example incorporating indirect information from other agents in the system [8, 14], or utilising social and cognitive concepts as part of the computation process [4]. Trust and reputation systems can also differ in the way they select a delegatee, for example by using the rating to weigh the likelihood of selection. While trust and reputation systems seek to satisfy many properties including resistance to different types of attacks by malicious agents [7], at their heart, they balance the exploration of delegatee behaviour with the exploitation of high quality delagatees.

The rapid advancement of chat-based language models has led to remarkable progress in complex task-solving. However, their success heavily relies on human input to guide the conversation, which can be challenging and time-consuming. This paper explores the potential of building scalable techniques to facilitate autonomous cooperation among communicative agents, and provides insight into their "cognitive" processes. To address the challenges of achieving autonomous cooperation, we propose a novel communicative agent framework named role-playing. Our approach involves using inception prompting to guide chat agents toward task completion while maintaining consistency with human intentions. We showcase how role-playing can be used to generate conversational data for studying the behaviors and capabilities of a society of agents, providing a valuable resource for investigating conversational language models. In particular, we conduct comprehensive studies on instruction-following cooperation in multi-agent settings. Our contributions include introducing a novel communicative agent framework, offering a scalable approach for studying the cooperative behaviors and capabilities of multi-agent systems, and open-sourcing our library to support research on communicative agents and beyond: https://github.com/camel-ai/camel.

This paper investigates how interacting agents arrive to a consensus or a polarized state. We study the opinion formation process under the effect of a global steering mechanism (GSM), which aggregates the opinion-driven stochastic agent states at the network level and feeds back to them a form of global information. We also propose a new two-layer agent-based opinion formation model, called GSM-DeGroot, that captures the coupled dynamics between agent-to-agent local interactions and the GSM's steering effect. This way, agents are subject to the effects of a DeGroot-like local opinion propagation, as well as to a wide variety of possible aggregated information that can affect their opinions, such as trending news feeds, press coverage, polls, elections, etc. Contrary to the standard DeGroot model, our model allows polarization to emerge by letting agents react to the global information in a stubborn differential way. Moreover, the introduced stochastic agent states produce event stream dynamics that can fit to real event data. We explore numerically the model dynamics to find regimes of qualitatively different behavior. We also challenge our model by fitting it to the dynamics of real topics that attracted the public attention and were recorded on Twitter. Our experiments show that the proposed model holds explanatory power, as it evidently captures real opinion formation dynamics via a relatively small set of interpretable parameters.

The ability to accurately predict others' behavior is central to the safety and efficiency of interactive robotics. Unfortunately, robots often lack access to key information on which these predictions may hinge, such as other agents' goals, attention, and willingness to cooperate. Dual control theory addresses this challenge by treating unknown parameters of a predictive model as stochastic hidden states and inferring their values at runtime using information gathered during system operation. While able to optimally and automatically trade off exploration and exploitation, dual control is computationally intractable for general interactive motion planning. In this paper, we present a novel algorithmic approach to enable active uncertainty reduction for interactive motion planning based on the implicit dual control paradigm. Our approach relies on sampling-based approximation of stochastic dynamic programming, leading to a model predictive control problem that can be readily solved by real-time gradient-based optimization methods. The resulting policy is shown to preserve the dual control effect for a broad class of predictive models with both continuous and categorical uncertainty. To ensure the safe operation of the interacting agents, we use a runtime safety filter (also referred to as a "shielding" scheme), which overrides the robot's dual control policy with a safety fallback strategy when a safety-critical event is imminent. We then augment the dual control framework with an improved variant of the recently proposed shielding-aware robust planning scheme, which proactively balances the nominal planning performance with the risk of high-cost emergency maneuvers triggered by low-probability agent behaviors. We demonstrate the efficacy of our approach with both simulated driving studies and hardware experiments using 1/10 scale autonomous vehicles.

In this paper, we focus on the data-driven discovery of a general second-order particle-based model that contains many state-of-the-art models for modeling the aggregation and collective behavior of interacting agents of similar size and body type. This model takes the form of a high-dimensional system of ordinary differential equations parameterized by two interaction kernels that appraise the alignment of positions and velocities. We propose a Gaussian Process-based approach to this problem, where the unknown model parameters are marginalized by using two independent Gaussian Process (GP) priors on latent interaction kernels constrained to dynamics and observational data. This results in a nonparametric model for interacting dynamical systems that accounts for uncertainty quantification. We also develop acceleration techniques to improve scalability. Moreover, we perform a theoretical analysis to interpret the methodology and investigate the conditions under which the kernels can be recovered. We demonstrate the effectiveness of the proposed approach on various prototype systems, including the selection of the order of the systems and the types of interactions. In particular, we present applications to modeling two real-world fish motion datasets that display flocking and milling patterns up to 248 dimensions. Despite the use of small data sets, the GP-based approach learns an effective representation of the nonlinear dynamics in these spaces and outperforms competitor methods.

Robust multi-agent reinforcement learning (MARL) necessitates resilience to uncertain or worst-case actions by unknown allies. Existing max-min optimization techniques in robust MARL seek to enhance resilience by training agents against worst-case adversaries, but this becomes intractable as the number of agents grows, leading to exponentially increasing worst-case scenarios. Attempts to simplify this complexity often yield overly pessimistic policies, inadequate robustness across scenarios and high computational demands. Unlike these approaches, humans naturally learn adaptive and resilient behaviors without the necessity of preparing for every conceivable worst-case scenario. Motivated by this, we propose MIR2, which trains policy in routine scenarios and minimize Mutual Information as Robust Regularization. Theoretically, we frame robustness as an inference problem and prove that minimizing mutual information between histories and actions implicitly maximizes a lower bound on robustness under certain assumptions. Further analysis reveals that our proposed approach prevents agents from overreacting to others through an information bottleneck and aligns the policy with a robust action prior. Empirically, our MIR2 displays even greater resilience against worst-case adversaries than max-min optimization in StarCraft II, Multi-agent Mujoco and rendezvous. Our superiority is consistent when deployed in challenging real-world robot swarm control scenario. See code and demo videos in Supplementary Materials.

A Multi-Agent Cooperative Learning (MACL) system is an artificial intelligence (AI) system where multiple learning agents work together to complete a common task. Recent empirical success of MACL systems in various domains (e.g. traffic control, cloud computing, robotics) has sparked active research into the design and analysis of MACL systems for sequential decision making problems. One important metric of the learning algorithm for decision making problems is its regret, i.e. the difference between the highest achievable reward and the actual reward that the algorithm gains. The design and development of a MACL system with low-regret learning algorithms can create huge economic values. In this thesis, I analyze MACL systems for different sequential decision making problems. Concretely, the Chapter 3 and 4 investigate the cooperative multi-agent multi-armed bandit problems, with full-information or bandit feedback, in which multiple learning agents can exchange their information through a communication network and the agents can only observe the rewards of the actions they choose. Chapter 5 considers the communication-regret trade-off for online convex optimization in the distributed setting. Chapter 6 discusses how to form high-productive teams for agents based on their unknown but fixed types using adaptive incremental matchings. For the above problems, I present the regret lower bounds for feasible learning algorithms and provide the efficient algorithms to achieve this bound. The regret bounds I present in Chapter 3, 4 and 5 quantify how the regret depends on the connectivity of the communication network and the communication delay, thus giving useful guidance on design of the communication protocol in MACL systems