One of the basic motivations for solving DCOPs is maintaining agents' privacy. Thus, researchers have evaluated the privacy loss of DCOP algorithms and defined corresponding notions of privacy preservation for secured DCOP algorithms. However, no secured protocol was proposed for Max-Sum, which is among the most studied DCOP algorithms. As part of the ongoing effort of designing secure DCOP algorithms, we propose P-Max-Sum, the first private algorithm that is based on Max-Sum. The proposed algorithm has multiple agents preforming the role of each node in the factor graph, on which the Max-Sum algorithm operates. P-Max-Sum preserves three types of privacy: topology privacy, constraint privacy, and assignment/decision privacy. By allowing a single call to a trusted coordinator, P-Max-Sum also preserves agent privacy. The two main cryptographic means that enable this privacy preservation are secret sharing and homomorphic encryption. In addition, we design privacy-preserving implementations of four variants of Max-Sum. We conclude by analyzing the price of privacy in terns of runtime overhead, both theoretically and by extensive experimentation.

Many real-world tasks involve multiple agents with partial observability and limited communication. Learning is challenging in these settings due to local viewpoints of agents, which perceive the world as non-stationary due to concurrently-exploring teammates. Approaches that learn specialized policies for individual tasks face problems when applied to the real world: not only do agents have to learn and store distinct policies for each task, but in practice identities of tasks are often non-observable, making these approaches inapplicable. This paper formalizes and addresses the problem of multi-task multi-agent reinforcement learning under partial observability. We introduce a decentralized single-task learning approach that is robust to concurrent interactions of teammates, and present an approach for distilling single-task policies into a unified policy that performs well across multiple related tasks, without explicit provision of task identity.

Multi-agent forward search (MAFS) is a state-of-the-art privacy-preserving planning algorithm. We describe a new variant of MAFS, called multi-agent forward-backward search (MAFBS) that uses both forward and backward messages to reduce the number of messages sent and obtain new privacy properties. While MAFS requires agents to send a state s produced by an action a to all agents that can apply any action in s, MAFBS sends such messages forward only to agents that have an action that requires one of the effects of a. To achieve completeness, it sends messages backward to agents that can supply a missing precondition. This more focused message passing scheme reduces states exchanged, and requires that agents be aware only of other agents that they directly interact with, leading to agent privacy.

Agents operating in a multi-agent environment must consider not just their own actions, but also those of the other agents in the system. Artificial social systems are a well known means for coordinating a set of agents, without requiring centralized planning or online negotiation between agents. Artificial social systems enact a social law which restricts the agents from performing some actions under some circumstances. A good social law prevents the agents from interfering with each other, but does not prevent them from achieving their goals. However, designing good social laws, or even checking whether a proposed social law is good, are hard questions. In this paper, we take a first step towards automating these processes, by formulating criteria for good social laws in a multi-agent planning framework. We then describe an automated technique for verifying if a proposed social law meets these criteria, based on a compilation to classical planning.

Swarm robotics has experienced a rapid expansion in recent years, primarily fueled by specialized multi-robot systems developed to achieve dedicated collective actions. These specialized platforms are in general designed with swarming considerations at the front and center. Key hardware and software elements required for swarming are often deeply embedded and integrated with the particular system. However, given the noticeable increase in the number of low-cost mobile robots readily available, practitioners and hobbyists may start considering to assemble full-fledged swarms by minimally retrofitting such mobile platforms with a swarm-enabling technology. Here, we report one possible embodiment of such a technology designed to enable the assembly and the study of swarming in a range of general-purpose robotic systems. This is achieved by combining a modular and transferable software toolbox with a hardware suite composed of a collection of low-cost and off-the-shelf components. The developed technology can be ported to a relatively vast range of robotic platforms with minimal changes and high levels of scalability. This swarm-enabling technology has successfully been implemented on two distinct distributed multi-robot systems, a swarm of mobile marine buoys and a team of commercial terrestrial robots. We have tested the effectiveness of both of these distributed robotic systems in performing collective exploration and search scenarios, as well as other classical cooperative behaviors. Experimental results on different swarm behaviors are reported for the two platforms in uncontrolled environments and without any supporting infrastructure. The design of the associated software library allows for a seamless switch to other cooperative behaviors, and also offers the possibility to simulate newly designed collective behaviors prior to their implementation onto the platforms.

The Belief-Desire-Intention (BDI) architecture is a practical approach for modelling large-scale intelligent systems. In the BDI setting, a complex system is represented as a network of interacting agents - or components - each one modelled based on its beliefs, desires and intentions. However, current BDI implementations are not well-suited for modelling more realistic intelligent systems which operate in environments pervaded by different types of uncertainty. Furthermore, existing approaches for dealing with uncertainty typically do not offer syntactical or tractable ways of reasoning about uncertainty. This complicates their integration with BDI implementations, which heavily rely on fast and reactive decisions. In this paper, we advance the state-of-the-art w.r.t. handling different types of uncertainty in BDI agents. The contributions of this paper are, first, a new way of modelling the beliefs of an agent as a set of epistemic states. Each epistemic state can use a distinct underlying uncertainty theory and revision strategy, and commensurability between epistemic states is achieved through a stratification approach. Second, we present a novel syntactic approach to revising beliefs given unreliable input. We prove that this syntactic approach agrees with the semantic definition, and we identify expressive fragments that are particularly useful for resource-bounded agents. Third, we introduce full operational semantics that extend CAN, a popular semantics for BDI, to establish how reasoning about uncertainty can be tightly integrated into the BDI framework. Fourth, we provide comprehensive experimental results to highlight the usefulness and feasibility of our approach, and explain how the generic epistemic state can be instantiated into various representations.

Inverse reinforcement learning (IRL) has become a useful tool for learning behavioral models from demonstration data. However, IRL remains mostly unexplored for multi-agent systems. In this paper, we show how the principle of IRL can be extended to homogeneous large-scale problems, inspired by the collective swarming behavior of natural systems. In particular, we make the following contributions to the field: 1) We introduce the swarMDP framework, a sub-class of decentralized partially observable Markov decision processes endowed with a swarm characterization. 2) Exploiting the inherent homogeneity of this framework, we reduce the resulting multi-agent IRL problem to a single-agent one by proving that the agent-specific value functions in this model coincide. 3) To solve the corresponding control problem, we propose a novel heterogeneous learning scheme that is particularly tailored to the swarm setting. Results on two example systems demonstrate that our framework is able to produce meaningful local reward models from which we can replicate the observed global system dynamics.

Sharon Glotzer has made a number of career-shifting discoveries, each one the kind "that completely changes the way you look at the world," she said, "and causes you to say, 'Wow, I need to follow this.'" A theoretical soft condensed matter physicist by training who now heads a thriving 33-person research group spanning three departments at the University of Michigan in Ann Arbor, Glotzer uses computer simulations to study emergence--the phenomenon whereby simple objects give rise to surprising collective behaviors. "When flocks of starlings make these incredible patterns in the sky that look like they're not even real, the way they're changing constantly--people have been seeing those patterns since people were on the planet," she said. "But only recently have scientists started to ask the question, how do they do that? How are the birds communicating so that it seems like they're all following a blueprint?"

This paper studies a criteria-based mechanism for nurturing and enhancing agents' group-benefiting individual efforts whenever the agents are self-interested. The idea is that only those agents that meet the criteria get to benefit from the group effort, giving an incentive to contribute even when it is otherwise individually irrational. Specifically, the paper provides a comprehensive equilibrium analysis of a threshold-based criteria mechanism for the common cooperative information gathering application, where the criteria is set such that only those whose contribution to the group is above some pre-specified threshold can benefit from the contributions of others. The analysis results in a closed form solution for the strategies to be used in equilibrium and facilitates the numerical investigation of different model properties as well as a comparison to the dual mechanism according to only an agent whose contribution is below the specified threshold gets to benefit from the contributions of others. One important contribution enabled through the analysis provided is in showing that, counter-intuitively, for some settings the use of the above-threshold criteria is outperformed by the use of the below-threshold criteria as far as collective and individual performance is concerned.

Decentralized Markov Decision Process (Dec-MDP) provides a rich framework to represent cooperative decentralized and stochastic planning problems under transition uncertainty. However, solving a Dec-MDP to generate coordinated yet decentralized policies is NEXP-Hard. Researchers have made significant progress in providing approximate approaches to improve scalability with respect to number of agents. However, there has been little or no research devoted to finding guarantees on solution quality for approximate approaches considering multiple (more than 2 agents) agents. We have a similar situation with respect to the competitive decentralized planning problem and the Stochastic Game (SG) model. To address this, we identify models in the cooperative and competitive case that rely on submodular rewards, where we show that existing approximate approaches can provide strong quality guarantees ( a priori, and for cooperative case also posteriori guarantees). We then provide solution approaches and demonstrate improved online guarantees on benchmark problems from the literature for the cooperative case.