In large-scale machine learning, recent works have studied the effects of compressing gradients in stochastic optimization in order to alleviate the communication bottleneck. These works have collectively revealed that stochastic gradient descent (SGD) is robust to structured perturbations such as quantization, sparsification, and delays. Perhaps surprisingly, despite the surge of interest in large-scale, multi-agent reinforcement learning, almost nothing is known about the analogous question: Are common reinforcement learning (RL) algorithms also robust to similar perturbations? In this paper, we investigate this question by studying a variant of the classical temporal difference (TD) learning algorithm with a perturbed update direction, where a general compression operator is used to model the perturbation. Our main technical contribution is to show that compressed TD algorithms, coupled with an error-feedback mechanism used widely in optimization, exhibit the same non-asymptotic theoretical guarantees as their SGD counterparts. We then extend our results significantly to nonlinear stochastic approximation algorithms and multi-agent settings. In particular, we prove that for multi-agent TD learning, one can achieve linear convergence speedups in the number of agents while communicating just $\tilde{O}(1)$ bits per agent at each time step. Our work is the first to provide finite-time results in RL that account for general compression operators and error-feedback in tandem with linear function approximation and Markovian sampling. Our analysis hinges on studying the drift of a novel Lyapunov function that captures the dynamics of a memory variable introduced by error feedback.

While multi-agent trust region algorithms have achieved great success empirically in solving coordination tasks, most of them, however, suffer from a non-stationarity problem since agents update their policies simultaneously. In contrast, a sequential scheme that updates policies agent-by-agent provides another perspective and shows strong performance. However, sample inefficiency and lack of monotonic improvement guarantees for each agent are still the two significant challenges for the sequential scheme. In this paper, we propose the \textbf{A}gent-by-\textbf{a}gent \textbf{P}olicy \textbf{O}ptimization (A2PO) algorithm to improve the sample efficiency and retain the guarantees of monotonic improvement for each agent during training. We justify the tightness of the monotonic improvement bound compared with other trust region algorithms. From the perspective of sequentially updating agents, we further consider the effect of agent updating order and extend the theory of non-stationarity into the sequential update scheme. To evaluate A2PO, we conduct a comprehensive empirical study on four benchmarks: StarCraftII, Multi-agent MuJoCo, Multi-agent Particle Environment, and Google Research Football full game scenarios. A2PO consistently outperforms strong baselines.

Robotics research has been focusing on cooperative multi-agent problems, where agents must work together and communicate to achieve a shared objective. To tackle this challenge, we explore imitation learning algorithms. These methods learn a controller by observing demonstrations of an expert, such as the behaviour of a centralised omniscient controller, which can perceive the entire environment, including the state and observations of all agents. Performing tasks with complete knowledge of the state of a system is relatively easy, but centralised solutions might not be feasible in real scenarios since agents do not have direct access to the state but only to their observations. To overcome this issue, we train end-to-end Neural Networks that take as input local observations obtained from an omniscient centralised controller, i.e., the agents' sensor readings and the communications received, producing as output the action to be performed and the communication to be transmitted. This study concentrates on two cooperative tasks using a distributed controller: distributing the robots evenly in space and colouring them based on their position relative to others. While an explicit exchange of messages between the agents is required to solve the second task, in the first one, a communication protocol is unnecessary, although it may increase performance. The experiments are run in Enki, a high-performance open-source simulator for planar robots, which provides collision detection and limited physics support for robots evolving on a flat surface. Moreover, it can simulate groups of robots hundreds of times faster than real-time. The results show how applying a communication strategy improves the performance of the distributed model, letting it decide which actions to take almost as precisely and quickly as the expert controller.

Recent work in the multi-agent domain has shown the promise of Graph Neural Networks (GNNs) to learn complex coordination strategies. However, most current approaches use minor variants of a Graph Convolutional Network (GCN), which applies a convolution to the communication graph formed by the multi-agent system. In this paper, we investigate whether the performance and generalization of GCNs can be improved upon. We introduce ModGNN, a decentralized framework which serves as a generalization of GCNs, providing more flexibility. To test our hypothesis, we evaluate an implementation of ModGNN against several baselines in the multi-agent flocking problem. We perform an ablation analysis to show that the most important component of our framework is one that does not exist in a GCN. By varying the number of agents, we also demonstrate that an application-agnostic implementation of ModGNN possesses an improved ability to generalize to new environments.

It plays a central role in intelligent agent systems to model agent's epistemic state and its change. Asynchrony plays a key role in distributed systems, in which the messages transmitted may not be received instantly by the agents. To characterize asynchronous communications, asynchronous announcement logic (AAL) has been presented, which focuses on the logic laws of the change of epistemic state after receiving information. However AAL does not involve the interactive behaviours between an agent and its environment. Through enriching the well-known pi-calculus by adding the operators for passing basic facts and applying the well-known action model logic to describe agents' epistemic states, this paper presents the e-calculus to model epistemic interactions between agents with epistemic states. The e-calculus can be adopted to characterize synchronous and asynchronous communications between agents. To capture the asynchrony, a buffer pools is constructed to store the basic facts announced and each agent reads these facts from this buffer pool in some order. Based on the transmission of link names, the e-calculus is able to realize reading from this buffer pool in different orders. This paper gives two examples: one is to read in the order in which the announced basic facts are sent (First-in-first-out, FIFO), and the other is in an arbitrary order.

We investigate the problem of autonomous racing among teams of cooperative agents that are subject to realistic racing rules. Our work extends previous research on hierarchical control in head-to-head autonomous racing by considering a generalized version of the problem while maintaining the two-level hierarchical control structure. A high-level tactical planner constructs a discrete game that encodes the complex rules using simplified dynamics to produce a sequence of target waypoints. The low-level path planner uses these waypoints as a reference trajectory and computes high-resolution control inputs by solving a simplified formulation of a racing game with a simplified representation of the realistic racing rules. We explore two approaches for the low-level path planner: training a multi-agent reinforcement learning (MARL) policy and solving a linear-quadratic Nash game (LQNG) approximation. We evaluate our controllers on simple and complex tracks against three baselines: an end-to-end MARL controller, a MARL controller tracking a fixed racing line, and an LQNG controller tracking a fixed racing line. Quantitative results show our hierarchical methods outperform the baselines in terms of race wins, overall team performance, and compliance with the rules. Qualitatively, we observe the hierarchical controllers mimic actions performed by expert human drivers such as coordinated overtaking, defending against multiple opponents, and long-term planning for delayed advantages.

Building ethical machines may involve bestowing upon them the emotional capacity to self-evaluate and repent on their actions. While reparative measures, such as apologies, are often considered as possible strategic interactions, the explicit evolution of the emotion of guilt as a behavioural phenotype is not yet well understood. Here, we study the co-evolution of social and non-social guilt of homogeneous or heterogeneous populations, including well-mixed, lattice and scale-free networks. Socially aware guilt comes at a cost, as it requires agents to make demanding efforts to observe and understand the internal state and behaviour of others, while non-social guilt only requires the awareness of the agents' own state and hence incurs no social cost. Those choosing to be non-social are however more sensitive to exploitation by other agents due to their social unawareness. Resorting to methods from evolutionary game theory, we study analytically, and through extensive numerical and agent-based simulations, whether and how such social and non-social guilt can evolve and deploy, depending on the underlying structure of the populations, or systems, of agents. The results show that, in both lattice and scale-free networks, emotional guilt prone strategies are dominant for a larger range of the guilt and social costs incurred, compared to the well-mixed population setting, leading therefore to significantly higher levels of cooperation for a wider range of the costs. In structured population settings, both social and non-social guilt can evolve and deploy through clustering with emotional prone strategies, allowing them to be protected from exploiters, especially in case of non-social (less costly) strategies. Overall, our findings provide important insights into the design and engineering of self-organised and distributed cooperative multi-agent systems.

Utilizing messages from teammates can improve coordination in cooperative Multi-agent Reinforcement Learning (MARL). Previous works typically combine raw messages of teammates with local information as inputs for policy. However, neglecting message aggregation poses significant inefficiency for policy learning. Motivated by recent advances in representation learning, we argue that efficient message aggregation is essential for good coordination in cooperative MARL. In this paper, we propose Multi-Agent communication via Self-supervised Information Aggregation (MASIA), where agents can aggregate the received messages into compact representations with high relevance to augment the local policy. Specifically, we design a permutation invariant message encoder to generate common information-aggregated representation from messages and optimize it via reconstructing and shooting future information in a self-supervised manner. Hence, each agent would utilize the most relevant parts of the aggregated representation for decision-making by a novel message extraction mechanism. Furthermore, considering the potential of offline learning for real-world applications, we build offline benchmarks for multi-agent communication, which is the first as we know. Empirical results demonstrate the superiority of our method in both online and offline settings. We also release the built offline benchmarks in this paper as a testbed for communication ability validation to facilitate further future research.

Deducing the contribution of each agent and assigning the corresponding reward to them is a crucial problem in cooperative Multi-Agent Reinforcement Learning (MARL). Previous studies try to resolve the issue through designing an intrinsic reward function, but the intrinsic reward is simply combined with the environment reward by summation in these studies, which makes the performance of their MARL framework unsatisfactory. We propose a novel method named Attention Individual Intrinsic Reward Mixing Network (AIIR-MIX) in MARL, and the contributions of AIIR-MIX are listed as follows:(a) we construct a novel intrinsic reward network based on the attention mechanism to make teamwork more effective. (b) we propose a Mixing network that is able to combine intrinsic and extrinsic rewards non-linearly and dynamically in response to changing conditions of the environment. We compare AIIR-MIX with many State-Of-The-Art (SOTA) MARL methods on battle games in StarCraft II. And the results demonstrate that AIIR-MIX performs admirably and can defeat the current advanced methods on average test win rate. To validate the effectiveness of AIIR-MIX, we conduct additional ablation studies. The results show that AIIR-MIX can dynamically assign each agent a real-time intrinsic reward in accordance with their actual contribution.

In multi-agent reinforcement learning (MARL), independent learners are those that do not observe the actions of other agents in the system. Due to the decentralization of information, it is challenging to design independent learners that drive play to equilibrium. This paper investigates the feasibility of using satisficing dynamics to guide independent learners to approximate equilibrium in stochastic games. For $\epsilon \geq 0$, an $\epsilon$-satisficing policy update rule is any rule that instructs the agent to not change its policy when it is $\epsilon$-best-responding to the policies of the remaining players; $\epsilon$-satisficing paths are defined to be sequences of joint policies obtained when each agent uses some $\epsilon$-satisficing policy update rule to select its next policy. We establish structural results on the existence of $\epsilon$-satisficing paths into $\epsilon$-equilibrium in both symmetric $N$-player games and general stochastic games with two players. We then present an independent learning algorithm for $N$-player symmetric games and give high probability guarantees of convergence to $\epsilon$-equilibrium under self-play. This guarantee is made using symmetry alone, leveraging the previously unexploited structure of $\epsilon$-satisficing paths.