Recent works on gradient-based attacks and universal perturbations can adversarially modify images to bring down the accuracy of state-of-the-art classification techniques based on deep neural networks to as low as 10% on popular datasets like MNIST and ImageNet. The design of general defense strategies against a wide range of such attacks remains a challenging problem. In this paper, we derive inspiration from recent advances in the fields of cybersecurity and multi-agent systems and propose to use the concept of Moving Target Defense (MTD) for increasing the robustness of a set of deep networks against such adversarial attacks. To this end, we formalize and exploit the notion of differential immunity of an ensemble of networks to specific attacks. To classify an input image, a trained network is picked from this set of networks by formulating the interaction between a Defender (who hosts the classification networks) and their (Legitimate and Malicious) Users as a repeated Bayesian Stackelberg Game (BSG).We empirically show that our approach, MTDeep reduces misclassification on perturbed images for MNIST and ImageNet datasets while maintaining high classification accuracy on legitimate test images. Lastly, we demonstrate that our framework can be used in conjunction with any existing defense mechanism to provide more resilience to adversarial attacks than those defense mechanisms by themselves.

Social dilemmas, where mutual cooperation can lead to high payoffs but participants face incentives to cheat, are ubiquitous in multi-agent interaction. We wish to construct agents that cooperate with pure cooperators, avoid exploitation by pure defectors, and incentivize cooperation from the rest. However, often the actions taken by a partner are (partially) unobserved or the consequences of individual actions are hard to predict. We show that in a large class of games good strategies can be constructed by conditioning one's behavior solely on outcomes (i.e.

The two common design principles for agent-based models, KISS (Keep It Simple, Stupid) and KIDS (Keep It Descriptive, Stupid) offer limited traction for developing cognitive agents, who typically have strong ties to research findings and established theories of cognition. A KIKS principle (Keep It Knowledgeable, Stupid) is proposed to capture the fact that cognitive agents are grounded in published research findings and theory, rather than simply selecting parameters in an ad-hoc way. In short, KIKS suggests that modelers should not focus on how many parameters, but should instead focus on choosing the right research papers and implement each of their key parameters and mechanisms. Based on this principle, a design process for creating cognitive agents based on cognitive models is proposed. This process is centered around steps that cognitive agent designers are already consider (e.g., literature search, validation, implementing a computational model).

In this article we describe a general cognitive model of human social behavior that is meant to increase the social intelligence of autonomous intelligent agents in different contexts. Despite the remarkable improvements that have been made on human-agent interaction, agents still have a limited capacity to be aware of the social reality that is present in the human mind and significantly guides human behavior. The model discussed in this paper is a step toward increasing that capacity significantly. Two different case studies are described in which the proposed model is used to better explain and predict human behavior. The first case study is the well known Ultimatum game. The second one is a variation of the "Game of Nines" played by children.

In this paper we explain how IETAL agents learn their environment, and how they build their intrinsic, internal representation of it, which they then use to build their expectations when on quest to satisfy its active drives. As environments change (with or without other agents present in them), the agents learn to new and "forget" irrelevant, "old" associations made. We discuss the concept of emotional context of associations, and show a gallery of simulations of behaviors in small multiagent societies.

In this paper we present a novel agent-based modeling methodology to predict rooftop solar adoptions in the residential energy market. We first applied several linear regression models to estimate missing variables for non-adopters, so that attributes of non-adopters and adopters could be used to train a logistic regression model. Then, we integrated the logistic regression model along with other predictive models into a multi-agent simulation platform and validated our models by comparing the forecast of aggregate adoptions in a typical zip code area with its ground truth. This result shows that the agent-based model can reliably predict future adoptions. Finally, based on the validated agent-based model, we compared the outcome of a hypothesized seeding policy with the original incentive plan, and investigated other alternative seeding policies which could lead to more adopters.

Developing collaborative robots that can productively operate out of isolation and work safely in uninstrumented, human-populated environments is critically important for advancing the field of robotics. Especially in domains where modern robots are ineffective, we wish to leverage human-robot teaming to improve the efficiency, ability, and safety of human workers. Our work, outlined in this extended abstract, focuses on creating agents capable of human-robot teamwork by leveraging learning from demonstration, hierarchical task networks, multi-agent planning and state estimation, and intention recognition. We briefly describe our recent work within human-robot collaboration, including task comprehension, learning and performing assistive behaviors, and training novice human collaborators to become competent co-workers.

This article describes the MultiAgent Decision Process (MADP) toolbox, a software library to support planning and learning for intelligent agents and multiagent systems in uncertain environments. Some of its key features are that it supports partially observable environments and stochastic transition models; has unified support for single- and multiagent systems; provides a large number of models for decision-theoretic decision making, including one-shot decision making (e.g., Bayesian games) and sequential decision making under various assumptions of observability and cooperation, such as Dec-POMDPs and POSGs; provides tools and parsers to quickly prototype new problems; provides an extensive range of planning and learning algorithms for single-and multiagent systems; and is written in C and designed to be extensible via the object-oriented paradigm.

We describe an implemented system that supports deep semantic NLU for controlling systems with multiple simulated robot agents. The system supports bidirectional communication for both human-agent and agent-agent inter-action. This interaction is achieved with the use of N-tuples, a novel form of Agent Communication Language using shared protocols with content expressing actions or intentions. The system's portability and flexibility is facilitated by its division into unchanging "core" and "application-specific" components.

Autonomous agents require interfaces to define their interactions with humans. The coupling between agents and humans is often limited, with disjoint goals between the agent interface and its associated autonomous components. This leads to a gap in human interaction relative to agent capabilities. We seek to aid interface designs by clarifying agent capabilities within an interface context. A taxonomy was developed that can help elucidate the agent's affordances and constraints that guide interface design. Moreover, the descriptors employed in the taxonomy can serve as a common language to support dialog between agent and interface developers, resulting in improved autonomous systems that support human-autonomy coordination.