Learning to Communicate: A Machine Learning Framework for Heterogeneous Multi-Agent Robotic Systems

This paper investigates a distributed time-varying output formation control framework for heterogeneous multi-agent systems (MASs) based on reinforcement learning (RL) method with multiple leaders and various disturbances. The outputs of followers are designed to accomplish robust tracking along with the movement of leaders’ convex combination while achieving a predefined configuration of time-varying formation. A three-layer framework, composed of a distributed adaptive finite-time observer, an off-policy RL-based optimal tracking controller and a robust formation controller, is proposed in a model-free manner while neither the global information of graph nor followers’ dynamics is utilized. The stability of this integrated observer-based controller is analyzed using Lyapunov theory which indicates that the formation tracking error asymptotically converges to zero under both external and internal unknown disturbances. Simulation results provide a detailed validation of the proposed control architecture.

In this paper, we develop a simple and efficient formation control framework for heterogeneous multi-agent systems under min-weighted persistent graph. As the ability of each agent may be different, the architecture of agents is considered to be heterogeneous. To reduce the communication complexity of keeping connectivity for agents, a topology optimisation scheme is proposed, which is based on min-weighted persistent graph. According to the topology of agents, a directed acyclic graph (DAG) is constructed to reflect the signal flow relation of agents, and then the corresponding formation control protocol is designed by using the transfer function model. Apply the proposed method, it is shown that the communication complexity of multi-agent systems is decreased, and the connection safety is improved. Based on signal flow graph analysis and Mason's rule, the convergence conditions are provided to show the agents can keep a formation. Finally, several simulations are worked out to illustrate the effectiveness of our theoretical results.

Motivated primarily by the problem of UAV coordination, in this paper we address the problem of coordination of a non-homogeneous group of non-holonomic agents with input constraints. In the first part of the paper, we develop a modeling framework for heterogeneous multi-agent systems that is based on timed automata. To this extent, an appropriate abstraction of the agents' workspace from our previous works is extended to three-dimensional space, by utilizing hexagonal prisms. The low level agent details are abstracted by virtue of appropriate controllers to motion primitives that can be performed in the individual workspace cells. The resulting models of the non-homogeneous system capture the non-holonomic behavior and the input constraints imposed by the considered systems. In the second part of this paper, we use the developed models in conjunction with formal verification tools to verify the safety and liveness properties of the system, captured by Linear Temporal Logic (LTL) specifications. Using counterexample guided search, we obtain trajectories that satisfy spatio-temporal specifications. Finally, we simulate two case-studies for two and three-dimensional workspaces respectively.

Reinforcement Learning has long been employed to solve sequential decision-making problems with minimal input data. However, the classical approach requires a long time to learn a suitable policy, especially in Multiagent Systems. The teacher-student framework proposes to mitigate this problem by integrating an advising procedure in the learning process, in which an experienced agent (human or not) can advise a student to guide her exploration. However, the teacher is assumed to be an expert in the learning task. We here propose an advising framework where multiple agents advise each other while learning in a shared environment, and the advisor is not expected to necessarily act optimally. Our experiments in a simulated Robot Soccer environment show that the learning process is improved by incorporating this kind of advice.

Emergence of game strategy in multiagent systems is studied. Symbolic and subsymbolic (neural network) approaches are compared. Symbolic approach is represented by a backtrack algorithm with specified search depth, whereas the subsymbolic approach is represented by feedforward neural networks that are adapted by reinforcement temporal difference TD(λ) technique. As a test game, we use simplified checkers. The problem is studied in the framework of multiagent system, where each agent is endowed with a neural network used for a classification of checkers positions. Three different strategies are used. The first strategy corresponds to a single agent that repeatedly plays games against MinMax version of a backtrack search method. The second strategy corresponds to single agents that are repeatedly playing a megatournament, where each agent plays two different games with all other agents, one game with white pieces and the other game with black pieces. After finishing each game, both agents modify their neural networks by reinforcement learning. The third strategy is an evolutionary modification of the second one. When a megatournament is finished, each agent is evaluated by a fitness, which reflects its success in the given megatournament (more successful agents have greater fitness). It is demonstrated that all these approaches lead to a population of agents very successfully playing checkers against a backtrack algorithm with the search depth 3.

This paper deals with the optimal bipartite consensus control (OBCC) for heterogeneous multi-agent systems (MASs) with time-delay based on reinforcement learning (RL) scheme. MASs, denoted by communication network, is modeled as a signed graph, and the competitive (cooperative) relationship is represented by negative (positive) edges. Firstly, Heterogeneous MASs is depicted by cooperative and competitive (hereinafter referred to as coopetition) network, and local neighborhood bipartite consensus errors and their performance index function (PIF) are put forward to formulate OBCC problem. Secondly, a proper model reduction method is developed to tackle time-delay in coopetition network, and then MASs with time-delay can be transformed into a delay-free MASs. Thirdly, a RL algorithm is employed to solve the solution of Hamilton-Jacobi-Bellman (HJB) equations under OBCC laws, and substantially we adopt neural networks (NNs) to calculate the control policy and PIF online. Finally, the pivotal simulation validates availability of proposed algorithm.

This paper investigates the error accumulation problem in centralized training and decentralized execution (CTDE) policy‐based multi‐agent reinforcement learning (MARL) algorithms, which arises from local observation inaccuracies. To address this issue, we propose a novel MARL algorithm that incorporates imitation learning using local observations. Firstly, by analysing the multi‐agent proximal policy optimization algorithm and examining the problems arising when global states are replaced with local observations, it is proved that insufficient observations can lead to information loss, thereby introducing errors of advantage function, and it is demonstrated that the generalized advantage estimation method accumulates errors during the training process. Then, imitation learning is introduced and a novel training framework that combines reinforcement learning and imitation learning is proposed. During the reinforcement learning phase, an MARL agent trained with global observations acts as an expert. Subsequently, imitation learning is applied to train another agent that mimics the expert's decisions using only local observations. Finally, the effectiveness of this algorithm is verified in some commonly used multi‐agent environments, which demonstrates its superior performance compared to traditional multi‐agent reinforcement learning algorithms.

In current multi-agent systems, the user is typically interacting with a single agent at a time through relatively inflexible and modestly intelligent interfaces. As a consequence, these systems force the users to submit simplistic requests only and suffer from problems such as the low-level nature of the system services offered to users, the weak reusability of agents, and the weak extensibility of the systems. In this paper, a framework for multi-agent systems called the open agent architecture (OAA) which reduces such problems, is discussed. The OAA is designed to handle complex requests that involve multiple agents. In some cases of complex requests from users, the components of the requests do not directly correspond to the capabilities of various application agents, and therefore, the system is required to translate the user's model of the task into the system's model before apportioning subtasks to the agents. To maximize users' efficiency in generating this type of complex requests, the OAA offers an intelligent multi-modal user interface agent which supports a natural language interface with a mix of spoken language, handwriting, and gesture. The effectiveness of the OAA environment including the intelligent distributed multi-modal interface has been observed in our development of several practical multi-agent systems.

Our overall aim is to provide a verification framework for practical multi-agent systems. To achieve practicality, we must be able to describe and implement heterogeneous multi-agent systems. To achieve verifiability, we must define semantics appropriately for use in formal verification. In this paper, we tackle the problem of implementing heterogeneous multi-agent systems in a semantically clear, and appropriate, way.

Reinforcement learning (RL) focuses on an essential aspect of intelligent behavior – how an agent can learn to make good decisions given experience and rewards in a stochastic<br>world. Yet popular RL algorithms that have enabled exciting successes in domains with good simulators (Go, Atari, etc) still often fail to learn in other domains because they rely on<br>simple heuristics for exploration. This provides additional empirical justification for essential questions around RL, specifically around algorithms that learn in a provably efficient manner through strategic exploration in any considered domain. This thesis provides new algorithms<br>and theory that enable good performance with respect to existing theoretical frameworks for evaluating RL algorithms (specifically, probably approximately correct) and introduces new stronger evaluation criteria, that may be particularly of interest as RL is applied to more real world problems.<br>For the first line of work on probably approximately correct (PAC) RL algorithms, we introduce a series of algorithms for episodic tabular domains with substantially better PAC<br>sample complexity bounds that culminate in a new algorithm with close to minimax optimal PAC and regret bounds. Look up tables are required by most sample efficient and computationally tractable algorithms, but cannot represent many practical domains. We therefore also present a new RL algorithm that can learn a good policy in environments with high dimensional observations and hidden deterministic states; unlike predecessors, this algorithm provably<br>explores not only in a statistically but also computationally efficient manner assuming access to function classes with efficient optimization oracles. To make progress it is critical to have the right measures of success. While empirical<br>demonstrations are quite clear, we find that for theoretical properties, two of the most commonly used learning frameworks, PAC guarantees and regret guarantees, each allow undesirable algorithm behavior (e.g. ignoring new observations that could improve the policy). We present<br>a new stronger learning framework called Uniform-PAC that unifies the existing frameworks and prevents undesirable algorithm properties. One caveat of all existing learning frameworks is that for any particular episode, we do not<br>know how well the algorithm will perform. To address this, we introduce the IPOC framework that requires algorithms to provide a certificate before each episode bounding how suboptimal the current policy can be. Such certifications may be of substantial interest in high stakes scenarios when an organization may wish to track or even pause an online RL system should the potential expected performance bound drop below a required expected outcome.

Daily streamflow forecasting by machine learning methods with weather and climate inputs

To guarantee the stability and performance of an overall heterogeneous multi-agent system, feedback control structures must be capable of addressing the presence of heterogeneous agent-based nonlinear uncertainties, unknown actuation performance, and coupling effects. Additionally, some tasks require agents to track the leader's state and reach different states in a distributed way. Therefore, this paper proposes a distributed adaptive controller that drives a set of agents to user-assigned positions in the presence of the mentioned heterogeneous system anomalies. To address actuator dynamics, a hedging-based reference model is used to decouple adaptive updating from the actuator dynamics. To deal with the coupled dynamics, an observer-based estimation algorithm is proposed. Finally, the low-frequency learning method is used to deal with the high-frequency control response and, therefore, to obtain a better transient performance. The overall closed-loop system's stability is investigated using the Lyapunov Stability Theory, and the Linear Matrix Inequalities (LMI) technique is employed to identify stability limits regarding the actuator bandwidths of each agent. Illustrative numerical examples are presented to demonstrate the designed controller's performance in a heterogeneous uncertain multi-agent system's dynamics with unknown actuation capabilities and coupled dynamics.

This paper gives aspects related to a cooperation scenario in the framework of multiagent systems. The presentation is focused on a multiagent system that consists of two agents, the Master and the Apprentice. The theoretical basis of the cooperation scenario is the definition of the most probable process, and two algorithms are used with this regard. The formulation of the cooperation scenario is exemplified for a case study that builds an architecture of successively placed bricks in the workspace.

Automatic layout generation of robotic production cells in a 3D manufacturing simulation environment

Reinforcement learning, a machine learning technique, has already shown advancement in solving complex sequential decision-making problems. Many tasks involve multiple agents and require sequential decision-making policies to achieve common goals, such as warehouse automation, autonomous driving, and game-playing. To derive the policies for all agents, these problems can be modelled as multi-agent systems and addressed by multi-agent reinforcement learning (MARL). However, optimizing policies in multi-agent scenarios presents significant challenges due to the intricate behaviours of multiple agents and the non-stationary nature of the environments’ complex dynamics. Firstly, optimizing policies in multi-agent scenarios presents significant challenges due to the complexity of multi-agent behaviours, especially in partially observable environments. Also, the dynamic nature of agents’ behaviours and their interactions with other agents lead to changes in the environment’s states and agents’ observations over time, which is more complicated in open environments. Furthermore, the need to balance individual and collective objectives in certain real-world multi-agent environments also complicates the decision-making process. This doctoral thesis aims to tackle the following three fundamental multi-agent research problems and propose a solution for each. Our research covers from theoretical analysis to practical application. We start by studying the problem of learning an efficient policy under partially observable environments. We notice that a group of agents cooperates to com- pete against another group of agents (named opponents), with the limitation that information about the opponents is inaccessible due to partial observability. To address this issue, we propose a novel multi-agent distributional actor-critic algorithm to achieve speculative opponent modeling with purely local information (i.e., the controlled agent’s observations, actions, and rewards). The actor maintains a speculated belief of the opponents, which we call the speculative opponent models, to predict opponents’ actions using local observations and make decisions accordingly. The distributional critic models the return distribution of the policy. It reflects the quality of the actor and thus can guide the training of the imaginary opponent model that the actor relies on. Extensive experiments confirm that our method successfully models opponents’ behaviours without their data and delivers superior performance against baseline methods with a faster convergence speed. Furthermore, in some environments, teammates’ numbers and policies change with the market, requiring workers to adapt to perform different sets of tasks across time. To solve this issue, we propose an RL-based method that allows the controlled agent to collaborate with dynamic teammates in open environments. The controlled agent maintains a dual teamwork situation inference model to capture the current teamwork state and facilitate reasonable decision-making under partial observability. Considering the dynamic types of teammates, we first leverage the Chinese Restaurant Process-based model to categorize versatile teammate policies into distinct clusters, which improves the efficiency of identifying current teamwork situations. Next, to model the heterogeneous relationships among agents and accommodate the varying number of teammates, we apply heterogeneous graph attention neural networks to learn the representation of the teamwork situation. Extensive experiments confirm that our method outperforms state-of-the-art baselines with faster convergence in various ad hoc teamwork tasks. Lastly, in certain real-world applications, such as routing problems and warehouse management, decision-makers must balance overall benefits with individual fairness among agents. Achieving both learning efficiency and fairness concurrently presents a complex, multi-objective, joint-policy optimization challenge. Moreover, existing approaches are predominantly confined to simulation environments. To address the above issues, we present a pioneering MARL approach to balance individual and collective objectives for agents collaboration. Extensive experiments on synthetic and real-world datasets demonstrate that our method not only surpasses the state-of-the-art DRL methods, but also exhibits advantages over established heuristics, notably in achieving significantly faster optimization speeds. This method underlines the essential integration of fairness into real-world applications, representing a major advancement in creating equitable and efficient logistics solutions. To conclude, this doctoral thesis investigates three fundamental multi-agent decision-making research problems that are ubiquitous and unsolved. The proposed three MARL methodology solutions achieve efficient policy training and performance for agents in multi-agent environments with uncertainties raised by the partially observable, the open environment, and the individual-collective objectives of MARL. This thesis delves into novel designs of various critical components of MARL, including MDP formulations, policy networks, training algorithms, and inference methods. These contributions significantly elevate the effectiveness and efficiency of cooperative MARL, establishing new performance benchmarks.