Dangers in Multiagent Rescue using DEFACTO

1
Dangers in Multiagent Rescue using DEFACTO

• Janusz Marecki
• Nathan Schurr, Milind Tambe,
• University of Southern California
• Paul Scerri
• Carnegie Mellon University

2
Dangers in Multiagent Rescue

• Autonomous Multiagent Rescue
• Problem Which house to rescue first?
• Human expertise responsibility
• Human supervisor
• Problem Human overwhelmed with tasks
• Mixed decision making DANGER

? ?
3
Outline

• Motivation and Domain
• DEFACTO System
• Adjustable Autonomy Strategies
• Predicted results
• Experimental results Dangers
• Summary

4
Motivation

• Large scale disasters
• Incident commander

5
Domain timeline

• Currently
• Thorough testing of DEFACTO system
• Short term goal
• Los Angeles Fire Department Training Tool
• Long term goal
• Automated First Responders
• under human supervision

6
Outline

• Motivation and Domain
• DEFACTO System
• Adjustable Autonomy Strategies
• Predicted results
• Experimental results Dangers
• Summary

7
DEFACTO System Architecture

• Demonstrating
• Effective
• Flexible
• Agent
• Coordination
• Through
• Omnipresence

8
DEFACTO System Architecture

• Robocup Rescue Simulation Environment
• 7 different simulators (fire, traffic, civilians
  etc.)
• Different maps (USC, Kobe)

• Demonstrating
• Effective
• Flexible
• Agent
• Coordination
• Through
• Omnipresence

9
DEFACTO System Architecture
10
DAFACTO Movie
11
DEFACTO System Architecture
Simulator
FireBrigade
FireBrigade
Machinetta Agent
Machinetta Agent
Machinetta Agent

• Machinetta Multiagent platform, Abstracted
  Theories of Teamwork (Scerri et al AAMAS 03)

12
Outline

• Motivation and Domain
• DEFACTO System
• Adjustable Autonomy Strategies
• Predicted results
• Experimental results Dangers
• Summary

13
Adjustable autonomy strategies

• 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, do not overload human with tasks!
• Previous Individual agent-human interaction

14
Team level Adjustable Autonomy

• AT Team level A strategy
• H Human strategy for all tasks
• AH Individual A strategy followed by
• the H strategy
• ATH Team level A strategy followed
• by the H strategy
• B The maximum number of agents the
• human is able to control
• EQH The quality of human decisions

15
Outline

• Motivation and Domain
• DEFACTO System
• Adjustable Autonomy Strategies
• Predicted results
• Experimental results Dangers
• Summary

16
Calculating predictions

• Strategy value equations
• Domain specific

17
Predicted results
Low B, Low EQh
Low B, High EQh

• Although higher expected quality of human
  decisions yields better results, low limit of
  human controllable agents hampers the overall
  score

18
Predicted results - ctnd
High B, Low EQh
High B, High EQh

• High limit of human controllable agents makes the
  human involving strategies effective also for
  larger teams, beating the fully autonomous A
  strategy

19
Outline

• Motivation and Domain
• DEFACTO System
• Adjustable Autonomy Strategies
• Predicted results
• Experimental results Dangers
• Summary

20
Experimental setup

• 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

21
Experimental results
22
Conclusions from results

• No strategy dominates through all the experiments
  in all cases
• As the number of agents increase, for strategy A
  the slope of improvement is greater than the
  slope of improvement for H. This correlates with
  our prediction that humans are not as good at
  exploiting additional agents resources, whereas
  agents are able to better exploit increasing
  numbers of available teammates
• If the difference for 4 agents between strategy A
  and H for a particular commander is small enough,
  as is the case with subjects A and C, then as we
  grow to larger numbers of agents, A will dominate
  AH, ATH and H
• ATH was constructed to help out at large of
  agents in the team. However, what we see instead
  is that ATH does better at smaller of agents
  over H, in a very surprising result. At higher
  of agents, ATH does worse for subject A than A.
• Dip at 6 agents?

23
Discrepancy for 6 agents?

• At 6 agents case, mixed strategies involving
  humans and agents (AH and ATH) performed worse
  than for 4 agents case
• At 6 agents case, H strategy improved over the 4
  agents case
• At 6 agents case, AT strategy improved over the 4
  agents case
• Hypothesis Human-Agent conflicts in resource
  allocation caused the problem

24
Task allocation overload danger
25
Summary

• Rigid transfer of control strategies are
  outperformed by flexible dominant strategy
  selection
• Having human in the loop does not necessary lead
  to increased performance
• Having humans and agents doing resource
  allocation simultaneously is susceptible to
  excessive reallocations which decreases overall
  performance

26
Future application

• Automated First Responders using DEFACTO

27
Thank you!

• Email marecki_at_usc.edu
• Teamcore web site http//teamcore.usc.edu
• Thanks
• CREATE Center
• Fred Pighin, Pratik Patil, Nikhil Kasinadhuni and
  J.P. Lewis