Transcript defacto
The DEFACTO System: Training Incident Commanders Nathan Schurr Janusz Marecki, Milind Tambe, Nikhil Kasinadhuni, and J. P. Lewis University of Southern California Paul Scerri Carnegie Mellon University Outline Motivation and Domain DEFACTO Team Level Adjustable Autonomy Experiments with DEFACTO Conclusions Motivation: Help Incident Commanders Incident Commander First Response Disaster Rescue Scenario Urban Environment Large Scale Crime Scene Incident commander must control situation, monitor situation, and allocate resources Goal: Initially a Training Simulation Later: Decision Support/Replacement LAFD Exercise: Simulations by People Playing Roles Aims of DEFACTO LAFD Exercise Challenges Key Exercise Components Personnel Heavy Smaller Scale Low Fidelity Environment Communication Allocation Agent-teams replace people playing roles Demonstrating Effective Flexible Agent Coordination of Teams via Omnipresence Outline Motivation and Domain DEFACTO Team Level Adjustable Autonomy Experiments with DEFACTO Conclusions DEFACTO Architecture Disaster Rescue Simulation: USC Map, Different underlying simulators Statistics Challenges in Extending to Human-Agent Teams Teamwork Communication Role Allocation Agent team to incorporate human Adjustable Autonomy (Scerri et al JAIR 2002) Interface DEFACTO Teamwork Proxies Machinetta Flexible Interaction Continued development with CMU Used in many other domains – UAVs, sensor nets etc. Team Level Adjustable Autonomy Strategies Dynamic Strategy Selection Omni-Viewer 2D – Standard with Simulator 3D – Developed by us Interaction DEFACTO Architecture Proxy Architecture Abstracted Theories of Teamwork (Scerri et al AAMAS 03) Communication: communication with other proxies Coordination: reasoning about team plans and communication State: the working memory of the proxy Adjustable Autonomy: reasoning about whether to act autonomously or pass control to the team member RAP Interface: communication with the team member Other Proxies Communication RAP Interface State Coordination Adjustable Autonomy RAP Teamwork Proxies Higher level TOP Reuse across domain Flexible Teamwork (Tambe JAIR 97) Communication Allocation Joint Intentions (Cohen & Levesque 1991) Role allocation algorithms (Xu et al AAMAS 2005) Machinetta Platform Independent Modular Structure Downloadable – Free, Publicly available Outline Motivation and Domain DEFACTO Team Level Adjustable Autonomy Experiments with DEFACTO Conclusions Adjustable Autonomy(AA) Strategies for Teams Agents dynamically adjust own level of autonomy Agents act autonomously, but also... Give up autonomy, transferring control to humans When to transfer decision-making control Whenever human has superior expertise Yet, too many interrupts also problematic Previous: Individual agent-human interaction AA: Novel Challenges in Teams Transfer of control strategies for AA in teams Planned sequence of transfers of control AT - Team level A strategy H - Human strategy for all tasks AH - Individual A followed by H ATH - Team level A strategy followed by H Goal: Improve Team Performance DEFACTO Architecture Omni-Viewer DEFACTO Movie Outline Motivation and Domain DEFACTO Team Level Adjustable Autonomy Experiments with DEFACTO Conclusions Experiments Initial evaluation of system and of strategies Details 3 Subjects Allocation Viewer Same Map for each scenario Building size and location Initial position of fires 4, 6, and 10 agents A, H, AH, ATH Strategies Averaged over 3 runs Empirical Studies with Users Subject B Subject A Buildings Saved 250 200 150 100 50 0 3 5 7 9 300 250 200 150 100 50 0 3 11 5 A H AH 7 A ATH H Subject C 300 250 200 150 100 50 0 3 5 9 Number of Fire Engines Number of Fire Engines Buildings Saved Buildings Saved 300 7 9 Number of Fire Engines A H AH ATH 11 AH ATH 11 Conclusions from Results No strategy dominates through all cases Humans may sometimes degrade agent team results Slope of strategy A > Slope of H Humans are not as good at exploiting additional agents resources If EQH is low, then as we grow to larger numbers of agents, A will dominate AH, ATH and H Dip at 6… LAFD – “Not surprising.” Summary DEFACTO Teamwork Team Level Adjustable Autonomy Strategies Interface Experimented with strategies for adjustable autonomy Future Directions Experiments with LAFD Study strategy behavior Train the “system” Training today, real response in the future. Future Thank You Email: [email protected] Web Site: http://teamcore.usc.edu Machinetta http://teamcore.usc.edu/doc/Machinetta/ Thanks CREATE Center Fred Pighin and Pratik Patil Related Work: Disaster Response Simulations LA County Fire Department Simulators DEFACTO focuses on “incident commander” “Environment” simulators: E.g., Terrasim, EPICS Not provide on agent behaviors “Agent-based” simulators E.g., Battlefield simulators Adjustable autonomy Strategy Models Models of the Strategies Models of the Strategies Outline Motivation DEFACTO Simulator Teamwork Proxies 3D Visualization Team Level Adjustable Autonomy Objectives CREATE Research Center Current State of the Art Models Predictions Experiments with DEFACTO Conclusions DEFACTO: Key Research Areas Enable effective interactions of agents with humans Scale-up to 100s of agents with fire engines, ambulances, police Research: Adjustable autonomy Previous work: Often single agent-single human interactions Research: Scale-up in team coordination Previous work: Limited numbers of agents coordinating in teams Visualization Robust 3D visualization Adjustable Autonomy: Novel Challenges in Teams Previous transfer-of-control fails in teams: Ignore costs to team (just concerned about individual) One shot transfers of control, too rigid Transfer control to a human (H) or agent (A) If human fails to make a decision, miscoordination!! Forcing agent to decide can cause a poor decision Expensive lesson learned in the “Electric-Elves” project Major errors by software assistants Hence need more flexible transfer of control Predictions EQh: Expected quality of human decision AGH: How many agents human can control A Strategy has constant slope Low B, High EQh Low B, Low EQh 80 Strategy value Strategy value 80 60 40 20 60 40 20 0 0 2 3 4 5 6 7 8 9 10 Number of agents A H ATH 2 3 4 5 6 7 8 9 10 Number of agents A H ATH CREATE Research Center Center for Risk and Economic Analysis of Terrorism Events MANPAD Scenario Large Scale Disaster Limited Resources First Response Help incident commander control situation Large Scale Crime Scene Simulator Robocup Rescue 10 different Simulators Multiple Agent Types Team Level AA Model How to select the strategy among many? Key idea: Calculate expected utility of different strategies Traditional Expected Utility Mathematical model of strategies EQ: Quality of an entity’s decision P: Probability of response of that entity W: Cost of miscoordination Probability of response * decision quality Integrate over time Agents Per Fire Subject B 4 3.5 3.5 Agents/Fire 4 3 2.5 3 2.5 2 3 4 5 6 7 8 9 Number of Agents AH ATH 10 11 2 3 4 5 6 7 8 Number of Agents AH Subject C 4 Agents/Fire Agents/Fire Subject A 3.5 3 2.5 2 3 4 5 6 7 8 9 Number of Agents AH ATH 10 11 9 ATH 10 11 LA City Fire Dept Exercise: Fire Progression Fire starts on 1st floor Spreads to Attic LAFD Exercise: Simulations by People Playing Roles LAFD officials simulate fire progression and the resource availability Battalion Chief allocates available resources to tasks Proxy Architecture Abstracted Theories of Teamwork (Machinetta) Platform Independent Modular Structure Other Proxies Communication RAP Interface State Coordination Adjustable Autonomy RAP DEFACTO Movie Objectives: Agent-based Simulation Tools for Disaster Response Improve training and decision making Present Future Teach and evaluate LAFD response tactics Agent/Robot disaster response Key research questions in: Multiagent coordination, Adjustable Autonomy Visualization of multiagent systems