Despite the potential for antisocial and counterproductive social media behavior–particularly in the context of humanitarian assistance/disaster response (HA/DR)–there is a paucity of automated methods to address it. Current methods focus primarily on detecting hate speech and banning problematic content. We propose an alternative strategy of using automated counter speech to focus not just on moderating uncivil behavior, but also the promotion of civil discourse. In this paper, we propose a novel framework to employ pre-trained language models to alleviate the bottlenecks in adoption of such counter speech, namely a lack of understanding on the dynamics of counter speech and a scarcity of well curated datasets, which are compounded in HA/DR settings. We utilize GPT LMs to create a conversational testbed to simulate online conversations where various approaches for counter speech and other content moderation methods can be evaluated. Additionally, we leverage BERT-based models to detect hate speech and other network and syntactic features to suggest the optimal strategy to employ. We also present empirical results on the experiments we have conducted which provide a proof of concept for the framework.

In recent years, advances in artificial intelligence (AI) have far outpaced our ability to understand and leverage them. In no domain has this been more true than in conversational agents (CAs). Transformer-based generative language models, such as GPT-2, significantly advance CAs’ ability to generate creative and relevant content. It is critical to start exploring collaboration with these CAs. In this paper, we focus on an initial step by enabling a human-augmented, AI-driven CA to contribute to a panel discussion. Key questions include training a transformer-based AI to talk like a panelist, effectively embodying the CA to interact with panel participants, and defining the operational requirements and challenges to a CA gaining acceptance from its peers. Our results highlight the benefits that varied training, equal and dynamic representation, and fluid operation can have for AI applications. While acknowledging limitations, we present a path forward to richer, more natural human-AI collaboration.

Today within the multiagent community, we see at least four competing methods to building multiagent systems: beliefdesire-intention (BDI), distributed constraint optimization (DCOP), distributed POMDPs, and auctions or game-theoretic methods. While there is exciting progress within each approach, there is a lack of cross-cutting research. This article highlights the various hybrid techniques for multiagent teamwork developed by the teamcore group. In particular, for the past decade, the TEAMCORE research group has focused on building agent teams in complex, dynamic domains. While our early work was inspired by BDI, we will present an overview of recent research that uses DCOPs and distributed POMDPs in building agent teams. While DCOP and distributed POMDP algorithms provide promising results, hybrid approaches allow us to use the complementary strengths of different techniques to create algorithms that perform better than either of their component algorithms alone. For example, in the BDI-POMDP hybrid approach, BDI team plans are exploited to improve POMDP tractability, and POMDPs improve BDI team plan performance.

Recently, cyber reasoning systems demonstrated near-human performance characteristics when they autonomously identified, proved, and mitigated vulnerabilities in software during a competitive event. New research seeks to augment human vulnerability research teams with cyber reasoning system teammates in collaborative work environments. However, the literature lacks a concrete understanding of vulnerability research workflows and practices, limiting designers’, engineers’, and researchers’ ability to successfully integrate these artificially intelligent entities into teams. This paper contributes a general workflow model of the vulnerability research process, and identifies specific collaboration challenges and opportunities anchored in this model. Contributions were derived from a qualitative field study of work habits, behaviors, and practices of human vulnerability research teams. These contributions will inform future work in the vulnerability research domain by establishing an empirically-driven workflow model that can be adapted to specific organizational and functional constraints placed on individual and teams.

Recent advances in artificial intelligence have demonstrated that the future of work will be defined by collaborative human-machine teams. In order to be effective, human-machine teams will rely on context-aware systems to enable collaboration. In this paper, we present three lessons learned from the past five years of developing context-aware systems that we believe will improve future system design. First, that semantic activity must captured, modeled, and analyzed to enable reasoning across missions, actors, and content. Second, that context-aware systems require multiple, federated data stores to optimize system and team performance. Finally, that real-time inter-actor communications are the essential feature enabling adaptation. We close with a discussion of the influences and implications that these lessons have on human-machine teaming, and outline future research activities that will be necessary before operationalizing these systems.

Future hybrid human-machine teams will need to optimize their performance in uncertain environments by adapting their team structure. To address this need, we have developed a framework based on minimization of variational free energy, an information theoretic measure that has been shown to account for a variety of biological self-organizing phenomena. This paper proposes a novel approach to balance team structure by adapting roles and relationships based upon this framework. We apply this approach to evaluate possible structures for an infantry squad of human soldiers and autonomous systems. Using our STATES team simulation environment, we simulate mission performance for these teams and demonstrate that this approach enables a 12-person team to achieve performance results on par with a 15-person traditional team in terms of mission execution time. We argue that these results indicate that the free energy approach will lead to better hybrid team adaptations and improved performance.

Increasingly advanced systems are becoming pervasive in society. They are taking on roles and have abilities that necessitate their teaming with humans. However, the authors argue that we will soon reach a point where the increased embedding of artificially intelligent systems into collaborative teams results in reduced performance. We believe that there is an upcoming "uncanny valley of collaboration" that must be traversed on the way to natural truly productive hybrid teams. In order to avoid potential obstacles to hybrid team progress and safety, we argue that this challenge should be clearly identified and addressed across the larger human system integration community. In this paper, we describe the problem, propose potential solutions, and provide some examples where we are making strides toward bridging this valley.

For humans and automation to collaborate and perform tasks effectively, all participants need access to a common representation of potentially relevant situational information, or context. This article describes a general framework for building context‐aware interactive intelligent systems that comprises three major functions: (1) capture human‐system interactions and infer implicit context; (2) analyze and predict user intent and goals; and (3) provide effective augmentation or mitigation strategies to improve performance, such as delivering timely, personalized information and recommendations, adjusting levels of automation, or adapting visualizations. Our goal is to develop an approach that enables humans to interact more intuitively and naturally with automation that is reusable across domains by modeling context and algorithms at a higher level of abstraction. We first provide an operational definition of context and discuss challenges and opportunities for exploiting context. We then describe our current work toward a general platform that supports developing context‐aware applications in a variety of domains. We then explore an example use case illustrating how our framework can facilitate personalized collaboration within an information management and decision support tool. Future work includes evaluating our framework.

Unmanned ground vehicles have the potential for supporting small dismounted teams in mapping facilities, maintaining security in cleared buildings, and extending the team’s reconnaissance and persistent surveillance capability. In order for such autonomous systems to integrate with the team, we must move beyond current interaction methods using heads-down teleoperation which require intensive human attention and affect the human operator’s ability to maintain local situational awareness and ensure their own safety. This paper focuses on the design, development and demonstration of a multimodal interaction system that incorporates naturalistic human gestures, voice commands, and a tablet interface. By providing multiple, partially redundant interaction modes, our system degrades gracefully in complex environments and enables the human operator to robustly select the most suitable interaction method given the situational demands. For instance, the human can silently use arm and hand gestures for commanding a team of robots when it is important to maintain stealth. The tablet interface provides an overhead situational map allowing waypoint-based navigation for multiple ground robots in beyond-line-of-sight conditions. Using lightweight, wearable motion sensing hardware either worn comfortably beneath the operator’s clothing or integrated within their uniform, our non-vision-based approach enables an accurate, continuous gesture recognition capability without line-of-sight constraints. To reduce the training necessary to operate the system, we designed the interactions around familiar arm and hand gestures.

One of the primary challenges facing the modern small-unit tactical team is the ability of the unit to safely and effectively search, explore, clear and hold urbanized terrain that includes buildings, streets, and subterranean dwellings. Buildings provide cover and concealment to an enemy and restrict the movement of forces while diminishing their ability to engage the adversary. The use of robots has significant potential to reduce the risk to tactical teams and dramatically force multiply the small unit's footprint. Despite advances in robotic mobility, sensing capabilities, and human-robot interaction, the use of robots in room clearing operations remains nascent. CHAMP is a software system in development that integrates with a team of robotic platforms to enable them to coordinate with a human operator performing a search and pursuit task. In this way, the human operator can either give control to the robots to search autonomously, or can retain control and direct the robots where needed. CHAMP's autonomy is built upon a combination of adversarial pursuit algorithms and dynamic function allocation strategies that maximize the team's resources. Multi-modal interaction with CHAMP is achieved using novel gesture-recognition based capabilities to reduce the need for heads-down tele-operation. The Champ Coordination Algorithm addresses dynamic and limited team sizes, generates a novel map of the area, and takes into account mission goals, user preferences and team roles. In this paper we show results from preliminary simulated experiments and find that the CHAMP system performs faster than traditional search and pursuit algorithms.

Intelligence, Surveillance, and Reconnaissance (ISR) sensing platforms are becoming increasingly complex. Consequently, the fidelity of collected data is continuing to increase, along with the number of deployable sensors that retrieve these data, such as those found on Remotely Piloted Aircraft (RPAs). There are numerous, critical challenges when designing ISR systems because the technology and the human are tightly integrated, resulting in interdependent performance and behaviors. Predicting operator error can inform more effective means of managing erroneous decisions, but current methods of doing so are impractical because of the effort required to construct operator models. We explored human performance in a target detection task by conducting a human-in-the-loop experiment that examined the performance of operators who simultaneously monitored four simulated RPA video feeds and determined the presence of targets at points of interest (POIs). The results of this experiment confirm that performance varies significantly across certain flight conditions (e.g., combinations of altitude, speed, aspect angle). A statistical model was constructed from the human data to predict operator error in new situations. In future work, the model predictions will be integrated with an automated flight planner that will adjust RPA air tasking orders in real time and intelligently revisit POIs when human error is likely.

Model-driven visualization (MDV) is a novel framework that supports more effective, intelligent user interfaces to improve decision making in complex environments by coupling cognitive and perceptual theories of information processing with advanced artificial intelligence methods. It embeds empirical and theory driven approaches for identifying and prioritizing data based on the information requirements and needs of the human decision maker within intelligent agents. The agents automatically deliver and present information based on its likely value using visualizations that best convey that information to the user(s) of the system. Agents also reason about the context and constraints of the user, environment, and display to enable a higher degree of personalization within an interactive user interface (e.g., by drawing a user's attention to interesting aspects of the data such as trends, anomalies, and patterns). We apply cognitive systems engineering processes to help identify the information available to individuals and/or teams, where it resides, where it is needed, and ultimately how to create the mappings required in connecting critical information to those who need it with innovative visualizations that most effectively support the end user. This paper describes the application of MDV to intelligently deliver timely, mission-critical information by adapting a Common Tactical Picture (CTP) display used for maritime situation awareness, threat assessment, and decision support.

The Association for the Advancement of Artificial Intelligence, in cooperation with Stanford University's Department of Computer Science, was pleased to present the 2012 Spring Symposium Series, held Monday through Wednesday, March 26–28, 2012, at Stanford University, Stanford, California, USA. The six symposia held were AI, the Fundamental Social Aggregation Challenge (cochaired by W. F. Lawless, Don Sofge, Mark Klein, and Laurent Chaudron); Designing Intelligent Robots (cochaired by George Konidaris, Byron Boots, Stephen Hart, Todd Hester, Sarah Osentoski, and David Wingate); Game Theory for Security, Sustainability, and Health (cochaired by Bo An and Manish Jain); Intelligent Web Services Meet Social Computing (cochaired by Tomas Vitvar, Harith Alani, and David Martin); Self-Tracking and Collective Intelligence for Personal Wellness (cochaired by Takashi Kido and Keiki Takadama); and Wisdom of the Crowd (cochaired by Caroline Pantofaru, Sonia Chernova, and Alex Sorokin). The papers of the six symposia were published in the AAAI technical report series.

The goal of representing context in a mixed initiative system is to model the information at a level of abstraction that is actionable for both the human and automated system. A potential solution to this problem is the Context for Human and Automation Teams (CHAT). This paper introduces the CHAT model and provides example implementations from several different applications such as task scheduling techniques, multi-agent systems, and human-robot interaction.

The Next Generation Air Transportation System will introduce new, advanced sensor technologies into the cockpit that must convey a large number of potentially complex alerts. Our work focuses on the challenges associated with prioritizing aircraft sensor alerts in a quick and efficient manner, essentially determining when and how to the pilot This alert decision becomes very difficult in NextGen due to the following challenges: 1) the increasing number of potential hazards, 2) the uncertainty associated with the state of potential hazards as well as pilot slate , and 3) the limited time to make safely-critical decisions. In this paper, we focus on pilot state and present a model for anticipating duration and quality of pilot behavior, for use in a larger system which issues aircraft alerts. We estimate pilot workload, which we model as being dependent on factors including mental effort, task demands. and task performance. We perform a mathematically rigorous analysis of the model and resulting alerting plans. We simulate the model in software and present simulated results with respect to manipulation of the pilot measures.

While NextGen operations are still under development, several key issues have already emerged, including increased information demands on the flight deck. The ALerting And Reasoning Management System (ALARMS) was designed as a strategic, automated system for combining and evaluating alert-related outputs from current and proposed NextGen systems. The model-driven interface integrates the status of the environment, pilot, and system to automatically present the most critical information at the right time, augmenting existing flight deck technologies. The current level of uncertainty in the environment and system as a whole is also evaluated and represented within the display. The four stage model of information processing presented by Parasuraman, Sheridan, & Wickens (2000) was used to guide the development of the underlying ALARMS automation. This document provides a brief overview of the ALARMS development process. Examples of the interface are included, and we discuss its implications.

In today's world, organizations are faced with increasingly large and complex problems that require decision-making under uncertainty. Current methods for optimizing such decisions fall short of handling the problem scale and time constraints. We argue that this is due to existing methods not exploiting the inherent structure of the organizations which solve these problems. We propose a new model called the OrgPOMDP (Organizational POMDP), which is based on the partially observable Markov decision process (POMDP). This new model combines two powerful representations for modeling large scale problems: hierarchical modeling and factored representations. In this paper we make three key contributions: (a) Introduce the OrgPOMDP model; (b) Present an algorithm to solve OrgPOMDP problems efficiently; and (c) Apply OrgPOMDPs to scenarios in an existing large organization, the Air and Space Operation Center (AOC). We conduct experiments and show that our Org-POMDP approach results in greater scalability and greatly reduced runtime. In fact, as the size of the problem increases, we soon reach a point at which the OrgPOMDP approach continues to provide solutions while traditional POMDP methods cannot. We also provide an empirical evaluation to highlight the benefits of an organization implementing an OrgPOMDP policy.