Large language models (LLMs) have revolutionized the field of artificial intelligence, endowing it with sophisticated language understanding and generation capabilities. However, when faced with more complex and interconnected tasks that demand a profound and iterative thought process, LLMs reveal their inherent limitations. Autonomous LLM-powered multi-agent systems represent a strategic response to these challenges. Such systems strive for autonomously tackling user-prompted goals by decomposing them into manageable tasks and orchestrating their execution and result synthesis through a collective of specialized intelligent agents. Equipped with LLM-powered reasoning capabilities, these agents harness the cognitive synergy of collaborating with their peers, enhanced by leveraging contextual resources such as tools and datasets. While these architectures hold promising potential in amplifying AI capabilities, striking the right balance between different levels of autonomy and alignment remains the crucial challenge for their effective operation. This paper proposes a comprehensive multi-dimensional taxonomy, engineered to analyze how autonomous LLM-powered multi-agent systems balance the dynamic interplay between autonomy and alignment across various aspects inherent to architectural viewpoints such as goal-driven task management, agent composition, multi-agent collaboration, and context interaction. It also includes a domain-ontology model specifying fundamental architectural concepts. Our taxonomy aims to empower researchers, engineers, and AI practitioners to systematically analyze the architectural dynamics and balancing strategies employed by these increasingly prevalent AI systems. The exploratory taxonomic classification of selected representative LLM-powered multi-agent systems illustrates its practical utility and reveals potential for future research and development.

Self-play (SP) is a popular multi-agent reinforcement learning (MARL) framework for solving competitive games, where each agent optimizes policy by treating others as part of the environment. Despite the empirical successes, the theoretical properties of SP-based methods are limited to two-player zero-sum games. However, for mixed cooperative-competitive games where agents on the same team need to cooperate with each other, we can show a simple counter-example where SP-based methods cannot converge to a global Nash equilibrium (NE) with high probability. Alternatively, Policy-Space Response Oracles (PSRO) is an iterative framework for learning NE, where the best responses w.r.t. previous policies are learned in each iteration. PSRO can be directly extended to mixed cooperative-competitive settings by jointly learning team best responses with all convergence properties unchanged. However, PSRO requires repeatedly training joint policies from scratch till convergence, which makes it hard to scale to complex games. In this work, we develop a novel algorithm, Fictitious Cross-Play (FXP), which inherits the benefits from both frameworks. FXP simultaneously trains an SP-based main policy and a counter population of best response policies. The main policy is trained by fictitious self-play and cross-play against the counter population, while the counter policies are trained as the best responses to the main policy's past versions. We validate our method in matrix games and show that FXP converges to global NEs while SP methods fail. We also conduct experiments in a gridworld domain, where FXP achieves higher Elo ratings and lower exploitabilities than baselines, and a more challenging football game, where FXP defeats SOTA models with over 94% win rate.

Multi-Agent Reinforcement Learning (MARL) has become a promising solution for constructing a multi-agent autonomous driving system (MADS) in complex and dense scenarios. But most methods consider agents acting selfishly, which leads to conflict behaviors. Some existing works incorporate the concept of social value orientation (SVO) to promote coordination, but they lack the knowledge of other agents' SVOs, resulting in conservative maneuvers. In this paper, we aim to tackle the mentioned problem by enabling the agents to understand other agents' SVOs. To accomplish this, we propose a two-stage system framework. Firstly, we train a policy by allowing the agents to share their ground truth SVOs to establish a coordinated traffic flow. Secondly, we develop a recognition network that estimates agents' SVOs and integrates it with the policy trained in the first stage. Experiments demonstrate that our developed method significantly improves the performance of the driving policy in MADS compared to two state-of-the-art MARL algorithms.

Fairness problems in recommender systems often have a complexity in practice that is not adequately captured in simplified research formulations. A social choice formulation of the fairness problem, operating within a multi-agent architecture of fairness concerns, offers a flexible and multi-aspect alternative to fairness-aware recommendation approaches. Leveraging social choice allows for increased generality and the possibility of tapping into well-studied social choice algorithms for resolving the tension between multiple, competing fairness concerns. This paper explores a range of options for choice mechanisms in multi-aspect fairness applications using both real and synthetic data and shows that different classes of choice and allocation mechanisms yield different but consistent fairness / accuracy tradeoffs. We also show that a multi-agent formulation offers flexibility in adapting to user population dynamics.

Reinforcement learning (RL) has shown promising results for real-time control systems, including the domain of plasma magnetic control. However, there are still significant drawbacks compared to traditional feedback control approaches for magnetic confinement. In this work, we address key drawbacks of the RL method; achieving higher control accuracy for desired plasma properties, reducing the steady-state error, and decreasing the required time to learn new tasks. We build on top of \cite{degrave2022magnetic}, and present algorithmic improvements to the agent architecture and training procedure. We present simulation results that show up to 65\% improvement in shape accuracy, achieve substantial reduction in the long-term bias of the plasma current, and additionally reduce the training time required to learn new tasks by a factor of 3 or more. We present new experiments using the upgraded RL-based controllers on the TCV tokamak, which validate the simulation results achieved, and point the way towards routinely achieving accurate discharges using the RL approach.

Predicting the future motion of traffic agents is vital for self-driving vehicles to ensure their safe operation. We introduce RedMotion, a transformer model for motion prediction that incorporates two types of redundancy reduction. The first type of redundancy reduction is induced by an internal transformer decoder and reduces a variable-sized set of road environment tokens, such as road graphs with agent data, to a fixed-sized embedding. The second type of redundancy reduction is a self-supervised learning objective and applies the redundancy reduction principle to embeddings generated from augmented views of road environments. Our experiments reveal that our representation learning approach can outperform PreTraM, Traj-MAE, and GraphDINO in a semi-supervised setting. Our RedMotion model achieves results that are competitive with those of Scene Transformer or MTR++. We provide an open source implementation that is accessible via GitHub and Colab. It is essential for self-driving vehicles to understand the relation between the motion of traffic agents and the surrounding road environment. Motion prediction aims to predict the future trajectory of traffic agents based on past trajectories and the given traffic scenario. Recent state-of-the-art methods (e.g., Shi et al. (2022); Wang et al. (2023); Nayakanti et al. (2023)) are deep learning methods trained using supervised learning.

This paper studies a decentralized multi-armed bandit problem in a multi-agent network. The problem is simultaneously solved by N agents assuming they face a common set of M arms and share the same arms' reward distributions. Each agent can receive information only from its neighbors, where the neighbor relationships among the agents are described by an undirected graph. Two fully decentralized multi-armed bandit algorithms are proposed, respectively based on the classic upper confidence bound (UCB) algorithm and the state-of-the-art KL-UCB algorithm. The proposed decentralized algorithms permit each agent in the network to achieve a better logarithmic asymptotic regret than their single-agent counterparts, provided that the agent has at least one neighbor, and the more neighbors an agent has, the better regret it will have, meaning that the sum is more than its component parts.

As modern systems become ever more connected with complex dynamic coupling relationships, the development of safe control methods for such networked systems becomes paramount. In this paper, we define a general networked model with coupled dynamics and local control and discuss the relationship of node-level safety definitions for individual agents with local neighborhood dynamics. We define a node-level barrier function (NBF), node-level control barrier function (NCBF), and collaborative node-level barrier function (cNCBF) and provide conditions under which sets defined by these functions will be forward invariant. We use collaborative node-level barrier functions to construct a novel distributed algorithm for the safe control of collaborating network agents and provide conditions under which the algorithm is guaranteed to converge to a viable set of safe control actions for all agents or a terminally infeasible state for at least one agent. We introduce the notion of non-compliance of network neighbors as a metric of robustness for collaborative safety for a given network state and chosen barrier function hyper-parameters. We illustrate these results on a networked susceptible-infected-susceptible (SIS) model.

This paper addresses the distributed online control problem over a network of linear time-invariant (LTI) systems (with possibly unknown dynamics) in the presence of adversarial perturbations. There exists a global network cost that is characterized by a time-varying convex function, which evolves in an adversarial manner and is sequentially and partially observed by local agents. The goal of each agent is to generate a control sequence that can compete with the best centralized control policy in hindsight, which has access to the global cost. This problem is formulated as a regret minimization. For the case of known dynamics, we propose a fully distributed disturbance feedback controller that guarantees a regret bound of $O(\sqrt{T}\log T)$, where $T$ is the time horizon. For the unknown dynamics case, we design a distributed explore-then-commit approach, where in the exploration phase all agents jointly learn the system dynamics, and in the learning phase our proposed control algorithm is applied using each agent system estimate. We establish a regret bound of $O(T^{2/3} \text{poly}(\log T))$ for this setting.

Robot swarms can be tasked with a variety of automated sensing and inspection applications in aerial, aquatic, and surface environments. In this paper, we study a simplified two-outcome surface inspection task. We task a group of robots to inspect and collectively classify a 2D surface section based on a binary pattern projected on the surface. We use a decentralized Bayesian decision-making algorithm and deploy a swarm of miniature 3-cm sized wheeled robots to inspect randomized black and white tiles of $1m\times 1m$. We first describe the model parameters that characterize our simulated environment, the robot swarm, and the inspection algorithm. We then employ a noise-resistant heuristic optimization scheme based on the Particle Swarm Optimization (PSO) using a fitness evaluation that combines decision accuracy and decision time. We use our fitness measure definition to asses the optimized parameters through 100 randomized simulations that vary surface pattern and initial robot poses. The optimized algorithm parameters show up to a 55% improvement in median of fitness evaluations against an empirically chosen parameter set.