A Model for Large scale Team Formation for Disaster Rescue Problem by mikeholy

VIEWS: 7 PAGES: 33

									  A Model for Large-
scale Team Formation
 for a Disaster Rescue
        Problem

       Balaji Viswanathan
       Graduate student in
       Computer Science
             UMBC
    Organization of this Talk
 What are Multi-agent systems?
 Problem Statement
 Team Formation – Background
 Motivation for the proposed Model
 Description of the model
 Design and Implementation
 Initial Results
 Discussion and Future Work
  Why do we need this field of
    Multi Agent Systems?
 Decentralization is the order of the day
 If societies could evolve to Democracy,
  Economies could evolve to Private
  ownership, Computer Architectures could
  evolve to distributed computing and
  Networks could evolve to Internet, WHY
  SHOULD AI BE LEFT BEHIND?
        Multi-Agents and AI
 Multi-Agent systems is the Artificial
  Intelligence’s answer to
  DISTRIBUTEDNESS
 If AI is analogous to psychology, Multi-
  Agent paradigm is analogous to Sociology
 Coordination, competition, collusion,
  cooperation and synchronization are
  required
               What are Agents?
 Agents attack the Intelligence problems in a
  distributed way
 Agents are Autonomous
 Agents are Self-interested
 Agents need to work with other agents in the
  society


 STATUTORY WARNING: ALL the Terms used in this slide are CONTROVERSIAL
  Application of Multi-Agents
 Integrating equipments in automated
  hospitals
 Defense – UAV coordination, robotic
  agents for peacekeeping, surveillance, etc.
 Rescue Applications
 Future Space exploration
 Modeling Economic and social systems
 Web based agents for intelligence in the
  web
The Problem – Disaster Rescue
         Domain constraints
 Domain can be extremely big and littered
  with large number of obstacles
 For rescue, the robots have to synchronize
  and cooperate
 The location of victims and obstacles are
  unknown, to start with
    Why do we need Multiple
           Agents?
 The domain can be millions of sq.m in size
  and a single robot can cover only 1sq.m
  per second.
 Cost and simplicity
 Robust and Fault tolerant
        Problem Statement
Given a huge and partially-visible
 environment, with large number of tasks
 that need to be synchronized, how do we
 build a model for forming teams among
 large number of heterogeneous agents?
           Team Formation
 Requires the joint cooperation for multiple
  agents, who can be self-interested
 Roles must be allocated and task sharing
  must be done
 There should be some modeling of the
  intentions of the other members of the
  team
           Existing work
 STEAM – Based on Joint Intentions
 Swarm Intelligence – Inspired from ant
  colonies
 Network models – Based on social model
  of networking
           Swarm Intelligence




Source: http://www.ieice.org/jpn/books/kaishikiji/200108/image/7.jpg
Social Networking
  Problem with Existing work


The approaches are not scalable to
 large scale multi-agent systems
Motivation for the proposed
          model
           Proposed Model
 Hierarchical control and communication
 Fluid Entities – Agents, Stationary Entities
  - Beacons
 Using decentralized mechanisms inspired
  from the natural systems – swarms of ants,
  bees, etc.
         Agents in the system
Usage of various task agents:
1. Explorer – Builds a map and deploys intelligent
   coordinators, called beacons
2. Searcher – Looks for victims in the domain
3. Confirmer – Accurate estimation of target
   position and task weight calculation
4. Rescuer – Which actually performs the rescue
Working - I
Working - II
Working - III
Working - IV
Working - V
Screenshot
   Parameters for Comparison
 Number of targets rescued
 Number of robots used per target
 Total distance traveled by the robots
 Communication cost
 Total size of the Target area that was
  covered
          Results – Number of targets
                    rescued
                  Time taken to rescue the targets with varying number of
                                          agents

                  60
Targets rescued




                  50                                                   50 Agents
                  40                                                   70 Agents
                  30                                                   100 Agents
                                                                       200 Agents
                  20
                                                                       500 Agents
                  10
                   0
                         00

                         00

                         00

                         00

                         00

                         00

                         00

                         00

                         00
                   0
                       10

                       20

                       30

                       40

                       50

                       60

                       70

                       80

                       90
                                       Time taken
                  Results – Targets rescued - II

                       Number of targets rescued with varying number of agents

                  60

                  50
Targets rescued




                  40

                  30

                  20

                  10

                  0
                        50          70        100         150        200         500
                                              Number of agents
                            Efficiency – Robots per target
                                    Efficiency in terms of targets rescued per agent

                             0.3
Targets rescued per agent




                            0.25

                             0.2
                                                                                             50
                            0.15
                                                                                             90
                             0.1

                            0.05

                              0
                                   50         100          200          300            500
                                                     Number of agents
                                Results – Distance Traveled
                                   Efficiency of task performance with respect to total distance
                                                             travelled

                               35000
Total Distance travelled per




                               30000
       target rescued




                               25000

                               20000

                               15000

                               10000

                                5000

                                   0
                                          50            80              100        200             500

                                                             Number of agents
                   Results - Communication
                       Time taken to rescue the targets with varying
                                   communication limits

                  60
                  50
Targets Rescued




                  40                                                         40
                  30                                                         50
                                                                             60
                  20
                                                                             100
                  10
                   0
                        00

                              00

                                    00

                                          00

                                                00

                                                      00

                                                            00

                                                                  00

                                                                        00
                   0
                       10

                             20

                                   30

                                         40

                                               50

                                                     60

                                                           70

                                                                 80

                                                                       90
                                              Time Taken
        Results – Target Area Covered
                           Time taken to rescue the targets with varying task domain
                                                     sizes

                  60

                  50
Targets Rescued




                  40
                                                                                           600 X 600
                                                                                           600 X 800
                  30
                                                                                           800 X 1000
                                                                                           1000 X 1000
                  20

                  10

                  0
                       0

                              00


                                     00


                                            00


                                                   00


                                                          00


                                                                 00


                                                                        00


                                                                               00


                                                                                      00
                            10


                                   20


                                          30


                                                 40


                                                        50


                                                               60


                                                                      70


                                                                             80


                                                                                    90
                                                    Time Taken
             Future Work
 Work further on learning
 Improve the efficiency
 Testing with greater domain areas, larger
  targets, larger agent teams
 Including other methodologies for
  comparison
               Questions
Enough of Listening!!!!

Now, IT’S your turn to Speak

								
To top