Besides facing the same challenges as single-agent systems, the distributed nature of complex multi-agent systems sparks many questions and problems revolving around the constraints imposed by communication. The idea that multi-agent systems require communication to access information, to coordinate or simply to sense the environment they are acting on is sometimes overlooked when thinking of (and solving) emerging theoretical challenges. However, research problems related to communication in Cyber-Physical Systems have been a prevalent target for network control research for decades. In particular, we take inspiration on Event Triggered Control to study how communication affects performance, safety and robustness in multi-agent systems. The work in this dissertation begins by covering a communication-based form of swarm robotics systems, where taking inspiration from ants, agents learn to forage cooperatively by communicating through the environment. We study what form of convergence guarantees we can derive in such systems and how these depend on the communication logic, proposing mean field formulations of such systems. We then draw an analogy between such learning-based swarms and distributed Reinforcement Learning (RL), and propose strategies to safely reduce communication of information in a general form of distributed Q-Learning problems. We extend these ideas to cooperative Multi-Agent RL systems where agents communicate state measurements with each-other, and define so-called robustness surrogate functions (value function robustness certificates). These certificates allow agents to distributedly estimate how robust the joint policies are against lack of information, and determine when do they need to update other agents with new measurements. At last, we look into the general problem of robust control in RL systems, and propose a characterization of policy robustness against state measurement noise that allows us to cast robustness as a secondary objective in a lexicographic optimization scheme, applicable to policy gradient algorithms. This answers the following premise: If we need to learn controllers that are then deployed in possibly uncertain environments, we may want to make sure that “robustifying” the controller does not decrease (excessively) the capacity of the controller to successfully solve the original problem (without uncertainty). The work presented through this dissertation covers different problems and jumps between overlapping fields, but the methods and techniques proposed share a common principle: As complex multi-agent systems become more applicable to engineering problems, the need for understanding (and simplifying) communication rules is increasingly motivated by safety. Therefore, the problems and solutions considered aim to advance towards a formal understanding and design of communication logic in complex, model free multi-agent systems. @en

We present a biologically inspired design for swarm foraging based on ant’s pheromone deployment, where the swarm is assumed to have very restricted capabilities. The robots do not require global or relative position measurements and the swarm is fully decentralized and needs no infrastructure in place. Additionally, the system only requires one-hop communication over the robot network, we do not make any assumptions about the connectivity of the communication graph and the transmission of information and computation is scalable versus the number of agents. This is done by letting the agents in the swarm act as foragers or as guiding agents (beacons). We present experimental results computed for a swarm of Elisa-3 robots on a simulator, and show how the swarm self-organizes to solve a foraging problem over an unknown environment, converging to trajectories around the shortest path, and test the approach on a real swarm of Elisa-3 robots. At last, we discuss the limitations of such a system and propose how the foraging efficiency can be increased.@en

Collaborative multiagent robotic systems, where agents coordinate by modifying a shared environment often result in undesired dynamical couplings that complicate the analysis and experiments when solving a specific problem or task. Simultaneously, biologically inspired robotics rely on simplifying agents and increasing their number to obtain more efficient solutions to such problems, drawing similarities with natural processes. In this work, we focus on the problem of a biologically inspired multiagent system solving collaborative foraging. We show how mean field techniques can be used to re-formulate such a stochastic multiagent problem into a deterministic autonomous system. This de-couples agent dynamics, enabling the computation of limit behaviors and the analysis of optimality guarantees. Furthermore, we analyse how having finite number of agents affects the performance when compared to the mean field limit and we discuss the implications of such limit approximations in this multiagent system, which have impact on more general collaborative stochastic problems.@en

We present an approach to safely reduce the communication required between agents in a Multi-Agent Reinforcement Learning system by exploiting the inherent robustness of the underlying Markov Decision Process. We compute robustness certificate functions (off-line), that give agents a conservative indication of how far their state measurements can deviate before they need to update other agents in the system with new measurements. This results in fully distributed decision functions, enabling agents to decide when it is necessary to communicate state variables. We derive bounds on the optimality of the resulting systems in terms of the discounted sum of rewards obtained, and show these bounds are a function of the design parameters. Additionally, we extend the results for the case where the robustness surrogate functions are learned from data, and present experimental results demonstrating a significant reduction in communication events between agents.@en

We present an approach to reduce the communication of information needed on a Distributed Q-Learning system inspired by Event Triggered Control (ETC) techniques. We consider a baseline scenario of a Distributed Q-Learning problem on a Markov Decision Process (MDP). Following an event-based approach, N agents sharing a value function explore the MDP and compute a trajectory-dependent triggering signal which they use distributedly to decide when to communicate information to a central learner in charge of computing updates on the action-value function. These decision functions form an Event Based distributed Q learning system (EBd-Q), and we derive convergence guarantees resulting from the reduction of communication. We then apply the proposed algorithm to a cooperative path planning problem, and show how the agents are able to learn optimal trajectories communicating a fraction of the information. Additionally, we discuss what effects (desired and undesired) these event-based approaches have on the learning processes studied, and how they can be applied to more complex multi-agent systems.@en

Ant Colony algorithms are a set of biologically inspired algorithms used commonly to solve distributed optimization problems. Convergence has been proven in the context of optimization processes, but these proofs are not applicable in the framework of robotic control. In order to use Ant Colony algorithms to control robotic swarms, we present in this work more general results that prove asymptotic convergence of a multi-agent Ant Colony swarm moving in a weighted graph.@en