As a paradigm for coordinating cooperative agents in dynamic environments, teamwork has been show... more As a paradigm for coordinating cooperative agents in dynamic environments, teamwork has been shown to be capable of leading to flexible and robust behavior. However, when teamwork is applied to the problem of building teams with hundreds of members, its previously existing, fundamental limitations become apparent. In this paper, we address the limitations of existing models as they apply to very large agent teams. We develop algorithms aimed at flexible and efficient coordination, applying a decentralized social network topology for team organization and the abstract coordination behaviors of Team Oriented Plans (TOPs). From this basis, we present a model to organize a team into dynamically evolving subteams, in order to flexibly coordinate the team. Additionally, we put forward a novel approach to sharing information within large teams, which provides for targeted, efficient information delivery with a localized reasoning process model built on previously incoming information. We have developed domain-independent software proxies, with which we demonstrate teams of an order of magnitude larger than those previously discussed in known published work. We implement the results of our approach, demonstrating its ability to handle the challenges of coordinating large agent teams.
IEEE Transactions on Cognitive and Developmental Systems, 2023
Learning interpretable communication is essential for multi-agent and human-agent teams (HATs). I... more Learning interpretable communication is essential for multi-agent and human-agent teams (HATs). In multi-agent reinforcement learning for partially-observable environments, agents may convey information to others via learned communication, allowing the team to complete its task. Inspired by human languages, recent works study discrete (using only a finite set of tokens) and sparse (communicating only at some time-steps) communication. However, the utility of such communication in human-agent team experiments has not yet been investigated. In this work, we analyze the efficacy of sparse-discrete methods for producing emergent communication that enables high agentonly and human-agent team performance. We develop agent-only teams that communicate sparsely via our scheme of Enforcers that sufficiently constrain communication to any budget. Our results show no loss or minimal loss of performance in benchmark environments and tasks. In human-agent teams tested in benchmark environments, where agents have been modeled using the Enforcers, we find that a prototype-based method produces meaningful discrete tokens that enable human partners to learn agent communication faster and better than a one-hot baseline. Additional HAT experiments show that an appropriate sparsity level lowers the cognitive load of humans when communicating with teams of agents and leads to superior team performance.
National Conference on Artificial Intelligence, Jul 13, 2008
Individual robots or agents will often need to form coalitions to accomplish shared tasks, e.g.,i... more Individual robots or agents will often need to form coalitions to accomplish shared tasks, e.g.,in sensor networks or markets. Furthermore, in most real systems it is infeasible for entities to interact with all peers. The presence of a social network can alleviate this problem by providing a neighborhood system within which entities interact with a reduced number of peers. Previous research has shown that the topology of the underlying social network has a dramatic effect on the quality of coalitions formed and consequently on system performance (Gaston & desJardins 2005a). It has also been shown that it is feasible to develop agents which dynamically alter connections to improve an organization's ability to form coalitions on the network. However those studies have not analysed the network topologies that result from connectivity adaptation strategies. In this paper the resulting network topologies were analysed and it was found that high performance and rapid convergence were attained because scale free networks were being formed. However it was observed that organizational performance is not impacted by limiting the number of links per agent to the total number of skills available within the population, implying that bandwidth was wasted by previous approaches. We used these observations to inform the design of a token based algorithm that attains higher performance using an order of magnitude less messages for both uniform and non-uniform distributions of skills.
Cognitive social simulations, enabled by cognitive architectures (such as ACT-R), are particularl... more Cognitive social simulations, enabled by cognitive architectures (such as ACT-R), are particularly well-suited for advancing our understanding of socially-distributed and sociallysituated cognition. As a result, multi-agent simulations featuring the use of ACT-R agents may be important in improving our understanding of the factors that influence collective sensemaking. While previous studies demonstrate the feasibility of using ACT-R to model collective cognition, as well as sensemaking processes at the individual level, the development of an ACT-R model of collective sensemaking in a coalition environment presents a range of relatively novel methodological, technological and modeling challenges. Such challenges include the need to equip ACT-R agents with communication capabilities, the need to deal with highly dynamic information environments, the need to support intelligent information retrieval capabilities, and the need to represent inter-agent cognitive differences. These challenges shape the nature of research and development efforts to create a multi-agent simulation capability that can be used to explore the impact of different sociotechnical interventions on collective sensemaking processes. In this paper, we discuss the research efforts being undertaken to address these challenges in the context of the International Technology Alliance (ITA) research program. We also discuss the motivations for using ACT-R to model collective sensemaking processes and outline some opportunities for model application and empirical evaluation.
Learning interpretable communication is essential for multi-agent and human-agent teams (HATs). I... more Learning interpretable communication is essential for multi-agent and human-agent teams (HATs). In multi-agent reinforcement learning for partially-observable environments, agents may convey information to others via learned communication, allowing the team to complete its task. Inspired by human languages, recent works study discrete (using only a finite set of tokens) and sparse (communicating only at some time-steps) communication. However, the utility of such communication in human-agent team experiments has not yet been investigated. In this work, we analyze the efficacy of sparse-discrete methods for producing emergent communication that enables high agentonly and human-agent team performance. We develop agent-only teams that communicate sparsely via our scheme of Enforcers that sufficiently constrain communication to any budget. Our results show no loss or minimal loss of performance in benchmark environments and tasks. In human-agent teams tested in benchmark environments, where agents have been modeled using the Enforcers, we find that a prototype-based method produces meaningful discrete tokens that enable human partners to learn agent communication faster and better than a one-hot baseline. Additional HAT experiments show that an appropriate sparsity level lowers the cognitive load of humans when communicating with teams of agents and leads to superior team performance.
Adaptive Agents and Multi-Agents Systems, May 2, 2011
Large heterogeneous teams in a variety of applications must make joint decisions using large volu... more Large heterogeneous teams in a variety of applications must make joint decisions using large volumes of noisy and uncertain data. Often not all team members have access to a sensor, relying instead on information shared by peers to make decisions. These sensors can become permanently corrupted through hardware failure or as a result of the actions of a malicious adversary. Previous work showed that when the trust between agents was tuned to a specific value the resulting dynamics of the system had a property called scale invariance which led to agents reaching highly accurate conclusion with little communication. In this paper we show that these dynamics also leave the system vulnerable to most agents coming to incorrect conclusions as a result of small amounts of anomalous information maliciously injected in the system. We conduct an analysis that shows that the efficiency of scale invariant dynamics is due to the fact that large number of agents can come to correct conclusions when the difference between the percentage of agents holding conflicting opinions is relatively small. Although this allows the system to come to correct conclusions quickly, it also means that it would be easy for an attacker with specific knowledge to tip the balance. We explore different methods for selecting which agents are Byzantine and when attacks are launched informed by the analysis. Our study reveals global system properties that can be used to predict when and where in the network the system is most vulnerable to attack. We use the results of this study to design an algorithm used by agents to effectively attack the network, informed by local estimates of the global properties revealed by our investigation.
In fields as diverse as sociology and physics researchers have been investigating the rich networ... more In fields as diverse as sociology and physics researchers have been investigating the rich networks that exist in nature. More recently, a small number of multi-agent researchers have shown that the performance of a group can be significantly impacted by the nature of the network that connects them. In this chapter, we build on these initial efforts, performing systematic experiments in an attempt to understand how and why social networks affect group performance. Our key conclusion is that performance of a team can sometimes be improved by imposing a social network with relatively few connections even if it were feasible to connect the agents with a complete network.
Friendship is very important. China has a proverb. "One chopstick is easily broken, but a dozen c... more Friendship is very important. China has a proverb. "One chopstick is easily broken, but a dozen can hardly be bent." -Matthew Polly, American Shaolin First and foremost, I have to thank my parents, Sudha and William Reese, for inspiring me with their intelligence, idealism and dedication. My Princeton friends-Stan Yamane, Matt Mullin, Erich Greene, and Dan Bernstein-have known me since my "technopeasant" days. I treasure all the time that I've spent playing Dungeons and Dragons with them; they taught me the meaning of teamwork and in many ways contributed to my research thinking. At Carnegie Mellon, there have been many people who have made my time at the university special. Due to my three year leave of absence at HP Labs, most of the women who entered the Ph.D. program with me-Ashley Stroupe, Bernardine Dias, Joelle Pineau, Rosemary Emery-Montemerlo, and Vandi Verma-have been gone for some years, but I fondly remember all the great potlucks we had together during my first two years in the program. Working with Women@SCS, in general, and Carol Frieze, in particular, has been a wonderful part of my CMU experience. Through the years, the RETSINA group has hosted many excellent researchers (too numerous to list) who have helped hone my research ideas. My office mate, Mary Berna-Koes, was a great source of support throughout the program. I've also been blessed with three talented research collaborators, Joseph Giampapa, Michael Mandel, and Reid van Lehn, who contributed to parts of the work described in this document. I am grateful to my advisor, Katia Sycara, for guiding me on my thesis journey and to my committee, Paul Scerri, Illah Nourbakhsh, and Milind Tambe, for their valuable insights. Mike Lewis provided useful feedback on the presentation of this material. In addition to having a world-class computer science program, Carnegie Mellon boasts some excellent martial arts training opportunities. This June, I received my black belt in Shotokan karate. I am grateful to all the people in the Pittsburgh Shotokan Karate clubmy instructors and fellow students-for helping me in my journey as a martial artist. But the person who has indisputably made the most difference in my life is my husband, Rahul Sukthankar, who is my best friend, my karate sparring partner, and my research inspiration. Without him, this thesis would not be possible.
Burn, W. Near-misses and future disaster preparedness. (conditionally accepted). Risk Analysis: A... more Burn, W. Near-misses and future disaster preparedness. (conditionally accepted). Risk Analysis: An International Journal. Dillon, R.L., Tinsley, C.H., & Burns, W. Evolving risk perceptions about near-miss terrorist events. (forthcoming). Decision Analysis Journal.
Proceedings of the fourth international joint conference on Autonomous agents and multiagent systems, 2005
Efficient coordination among large numbers of heterogeneous agents promises to revolutionize the ... more Efficient coordination among large numbers of heterogeneous agents promises to revolutionize the way in which some complex tasks, such as responding to urban disasters can be performed. However, state of the art coordination algorithms are not capable of achieving efficient and effective coordination when a team is very large. Building on recent successful token-based algorithms for task allocation and information sharing, we have developed an integrated and efficient approach to effective coordination of large scale teams. We use tokens to encapsulate anything that needs to be shared by the team, including information, tasks and resources. The tokens are efficiently routed through the team via the use of local decision theoretic models. Each token is used to improve the routing of other tokens leading to a dramatic performance improvement when the algorithms work together. We present results from an implementation of this approach which demonstrates its ability to coordinate large teams.
Incorporating coalition formation algorithms into agent systems shall be advantageous due to the ... more Incorporating coalition formation algorithms into agent systems shall be advantageous due to the consequent increase in the overall quality of task performance. Coalition formation was addressed in game theory, h o wever the game theoretic approach is centralized and computationally intractable. Recent work in DAI has resulted in distributed algorithms with computational tractability. This paper addresses the implementation of distributed coalition formation algorithms within a real-world multi-agent system. We present the problems that arise when attempting to utilize the theoretical coalition formation algorithms for a real-world system, demonstrate how some of their restrictive assumptions can be relaxed, and discuss the resulting bene ts. In addition, we analyze the modi cations, the complexity and the quality of the cooperation mechanisms. The task domain of our multi-agent system is information gathering, ltering and decision support within the WWW.
2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology, 2009
Principle: By equipping software agents with domain-independent models of teamwork, agents are 1)... more Principle: By equipping software agents with domain-independent models of teamwork, agents are 1) able to assume the role of full team members 2) more capable of recognizing, monitoring and aiding teamwork activities between human teammates. Agent serving as human team member assistant
2007 10th International Conference on Information Fusion, 2007
In today's fast paced military operational environment, vast amounts of information must be sorte... more In today's fast paced military operational environment, vast amounts of information must be sorted out and fused not only to allow commanders to make situation assessments, but also to support the generation of hypotheses about enemy force disposition and enemy intent. An automation methodology and support tools are required to allow commanders to model and assess dynamic situations such as the behavior and intentions of enemy forces based on the flow and fusion of collected information from various sensors. Battlefield situation awareness, including location, movement, and deployment of enemy forces, is essential for commanders to make better decisions than adversaries. Agent-based information fusion and dissemination has great promise to considerably enhance current fusion technology. This promise stems from the distributed nature of agent technology, its flexibility and robustness in the face of network and sensor failures and its promise of performing fusion at run time.
Case-Based Reasoning Research and Development, 2001
This paper describes a prototype in which a conversational case-based reasoner, NaCoDAE, was agen... more This paper describes a prototype in which a conversational case-based reasoner, NaCoDAE, was agenti ed and inserted in the RET-SINA multi-agent system. Its task was to determine agent roles within a heterogeneous society of agents, where the agents may use capabilitybased or team-oriented agent coordination strategies. There were three reasons for assigning this task to NaCoDAE: 1 to relieve the agents of the overhead of determining, for themselves, if they should be involved in the task, or not; 2 to convert seemingly unrelated data into contextually relevant knowledge | as a case-based reasoning system, NaCo-DAE is particularly suited for applying apparently incoherent data to a wide variety of domain-speci c situations; and 3 as a conversational CBR system, to both unobtrusively listen to human statements and to proactively dialogue with other agents in a more goal-directed approach to gathering relevant information. The cases maintained by NaCoDAE have question and answer components, which w ere originally intended to maintain the textual representations of questions and answers for humans. By associating agent capability descriptions and queries with the case questions, NaCoDAE also assumed the team role of a capabilitybased coordinator. By encoding fragments of HTN plan objectives in its case actions, we w ere able to convert NaCoDAE into a conversational case-based planner that served compositionally-generated HTN plan objectives, already populated with situation-relevant knowledge, for use by the RETSINA team-oriented agents. 1
Lecture Notes in Control and Information Sciences, 2009
The performance of a cooperative team depends on the views that individual team members build of ... more The performance of a cooperative team depends on the views that individual team members build of the environment in which they are operating. Teams with many vehicles and sensors generate a large amount of information from which to create those views. However, bandwidth limitations typically prevent exhaustive sharing of this information. As team size and information diversity grows, it becomes even harder to provide agents with needed information within bandwidth constraints, and it is impractical for members to maintain any detailed information for every team mate. Building on previous token-based algorithms, this chapter presents an approach for efficiently sharing information in large teams. The key distinction from previous work is that this approach models differences in how agents in the team value knowledge and certainty about features. By allowing the tokens passed through the network to passively estimate the value of certain types of information to regions of the network, it is possible to improve token routing through the use of local decision-theoretic models. We show that intelligent routing and stopping can increase the amount of locally useful information received by team members while making more efficient use of agents' communication resources.
2007 10th International Conference on Information Fusion, 2007
A cooperative team's performance strongly depends on the view that the team has of the environmen... more A cooperative team's performance strongly depends on the view that the team has of the environment in which it operates. In a team with many autonomous vehicles and many sensors, there is a large volume of information available from which to create that view. However, typically communication bandwidth limitations prevent all sensor readings being shared with all other team members. This paper presents three policies for sharing information in a large team that balance the value of information against communication costs. Analytical and empirical evidence of their effectiveness is provided. The results show that using some easily obtainable probabilistic information about the team dramatically improves overall belief sharing performance. Specifically, by collectively estimating the value of a piece of information, the team can make most efficient use of its communication resources.
As a paradigm for coordinating cooperative agents in dynamic environments, teamwork has been show... more As a paradigm for coordinating cooperative agents in dynamic environments, teamwork has been shown to be capable of leading to flexible and robust behavior. However, when teamwork is applied to the problem of building teams with hundreds of members, its previously existing, fundamental limitations become apparent. In this paper, we address the limitations of existing models as they apply to very large agent teams. We develop algorithms aimed at flexible and efficient coordination, applying a decentralized social network topology for team organization and the abstract coordination behaviors of Team Oriented Plans (TOPs). From this basis, we present a model to organize a team into dynamically evolving subteams, in order to flexibly coordinate the team. Additionally, we put forward a novel approach to sharing information within large teams, which provides for targeted, efficient information delivery with a localized reasoning process model built on previously incoming information. We have developed domain-independent software proxies, with which we demonstrate teams of an order of magnitude larger than those previously discussed in known published work. We implement the results of our approach, demonstrating its ability to handle the challenges of coordinating large agent teams.
The socially-distributed nature of cognitive processing in a variety of organizational settings m... more The socially-distributed nature of cognitive processing in a variety of organizational settings means that there is increasing scientific interest in the factors that affect collective cognition. In military coalitions, for example, there is a need to understand how factors such as communication network topology, trust, cultural differences and the potential for miscommunication affects the ability of distributed teams to generate high quality plans, to formulate effective decisions and to develop shared situation awareness. The current paper presents a computational model and associated simulation capability for performing in silico experimental analyses of collective sensemaking. This model can be used in combination with the results of human experimental studies in order to improve our understanding of the factors that influence collective sensemaking processes.
Uploads
Papers by Katia Sycara