Integrating Intelligent Assistants into Human Teams Katia Sycara - - PowerPoint PPT Presentation
Integrating Intelligent Assistants into Human Teams Katia Sycara Michael Lewis The Robotics Institute School of Information Sciences Carnegie Mellon University University of Pittsburgh Pittsburgh, PA 15213 Pittsburgh, PA 15260 (412)
Katia Sycara The Robotics Institute Carnegie Mellon University Pittsburgh, PA 15213 (412) 268-8825 katia@cmu.edu
www.cs.cmu.edu/˜softagents
Michael Lewis School of Information Sciences University of Pittsburgh Pittsburgh, PA 15260 (412) 624-9426 ml@sis.pitt.edu
www.pitt.edu/˜cmlewis
Prasad Chalasani Liren Chen Keith Decker Kostya Domashnev Somesh Jha Anadeep Pannu Onn Shehory Rande Shern Vandana Verma Dajun Zeng
Michael Lewis (PI) Terry Lenox Emily Roth
Increase the effectiveness of joint Command and Control Teams through the incorporation of Agent Technology in environments that are:
distributed time stressed uncertainappear and disappear) Team members are distributed in terms of:
time and space expertise RETSINA infrastructure) that interact with the humans and each other. – integrate multimedia information management and decision support – anticipate and satisfy human information processing and problem solving needs – perform real-time synchronization of human actions – notify about significant changes in the environment – adapt to user, task and situation
develop model libraries of individual and team tasks develop verifiable useful human-agent interaction techniques gathering in a multiagent setting
single agent architecture and self-awareness agent coordination scheme finding appropriate agents agent interoperability agent-to-agent task delegation protocols learning through agent interactions simulations and perform controlled experiments with human subjects to assess utility)
human-agent trust development of task models (graphical task editor) user-guided instantiation of agents (agent editor) feedback, summary of most important information in documents)
detecting “interesting” patterns from multiple data sources information integration and conflict resolution Service Request USER 1 USER 2 USER h query answer Conflict Resolution Information Integration Information Request Reply
Interface Agent 2 Interface Agent k InfoAgent 1 InfoAgent n
Task Proposed Solution Task
Interface Agent 1
Goals and Task Specifications Results
Info Source 1 Info Source k Info Source 2
TaskAgent j TaskAgent 2 TaskAgent 1
Advertisement
MiddleAgent 1
– Goal-directed – Taskable – Running unassisted for long periods – Proactive & Reactive
– Agents are taskable, i.e. users or other agents can delegate tasks to them, user acceptability and trust an important issue – Can interact as cooperative teams or self-interested individuals – Interaction protocols – Coordination Strategies – Negotiation Protocols
Agents adapt to their environment, user, task and each other– Adapt both at the individual level and at the societal level – Employ Alternate Methods – Learn from (and about) users and each other
– Java implementation
RETSINA agent architecture– increased planning sophistication in individual agents
Middle agents Agent interaction protocols preferences initially known by Capabilities initially known by provider
provider + mid- dle agent provider + middle + re- quester requester only (broadcaster) “front-agent” matchmaker/yellow- pages requester + mid- dle agent anonymizer broker recommender requester + mid- dle + provider blackboard introducer/- bodyguard arbitrator
Communications Monitor Execution Scheduler Planner and Control Flow Data Flow KQML Messages to & from
Control Knowledge
Current Action Action Domain Facts Plan Library Objectives Schedule Task Structures Beliefs Database
– alternative task reductions – contingent plans, loops
incremental task reduction, interleaved with execution– information gathered during execution directs future planning
resource and temporal constraints Advertise Provision Down Parameter Outcome Get Middle-Agent Name Do-It Send KQML Message Reply-With Receiver Content Make Advertisement OK OK OK
– language for capability advertisement (Aardvark) – agent name server and distributed matchmakinga
Human Agent Interaction– Task Editor – Agent Editor – Human Agent Trust – Team TANDEM experiments
awww.cs.cmu.edu/˜softagents/retsina/ans
– Information filtering: Webmatea, DVINA – Agents in team aiding: ModSAF , multiagent air patrol, agent-aided aircraft maintenanceb
awww.cs.cmu.edu/˜softagents/webmate bThis application is done in collaboration with the CMU wearable computer project.
ModSAF
Region II Region III Agent Doctrine WebMate Weather Materiel Case-Based Agent
Infosphere
Region I Agent Path Planning Agent 2 Agent 3 USER 1 USER 2 USER 3 Agent 1 Interface Interface Interface Agent Team Agent Analysis Force Agent Agent Agent Agent Agent Agent
domains which are automatically learned
Use the trigger pair model to automatically extract relevant words for refiningsearch
The user can provide multiple pages as relevance guidance for informationsearch
, DARPA, NASA and others in collaboration with WQED
A multimedia library that will consist of over one thousand hours of digitalvideo, audio, images, text and other related materials
Uses combined speech, language and image understanding technology totranscribe, segment and index the linear video.
sophisticated planning mechanisms)
Improve the robustness of the RETSINA infrastructure Finish the implementation of the agent advertisement language (Aardvark) Refine agent task delegation framework, particularly contingent taskdelegation
Investigate situation-dependent agent coordination strategies Investigate information- and action-based conflict resolution Expand the ModSAF team-aiding scenarios by introducing agents ofadditional types and functionalities
activity traces
for agent-based team aiding
ModSAF scenarios with human subjects
Incorporate multimedia information from Informedia into agent-based teamaiding
Use the wearable computers as the platform for running the collaborativeaircraft maintenance agents