Coalition Agents eXperiment The Coalition TIE

1


CoAX Stand-alone Contributions DARPA Briefing -
November 2000 Dartmouth College, UMichigan, MIT
Sloan, Coalition Agents eXperiment
(CoAX) http//www.aiai.ed.ac.uk/project/coax/
2
Stand-alone Contributions

• Dartmouth Field Observation Agent
• MIT Robustness Service
• Michigan Coordination Planning Aid

3
Field Observations (Dartmouth)

• ActComm Project
• Dartmouth, Harvard, RPI, Illinois, ALPHATECH,
  Lockheed Martin
• Department of Defense Multidisciplinary
  University Research Initiative
• Developing a system to provide network access to
  soldiers in the field
• CoAX Goal
• Demonstrate the ease with which the large ActComm
  legacy system can be integrated with the rest
  of CoAX via the DARPA CoABS Grid

4
Field Observations (Dartmouth)

• Team of soldiers
• PDAs
• Ad-hoc wireless networking
• Soldiers make observations.
• Ground and air traffic
• Personnel and equipment
• Buildings and other structures
• Observations fed into battle-planning systems
  (e.g., MBP) through the CoABS Grid.
• In the demo, a team of CoAX soldiers will make
  observations to correct Gao misinformation.

5
Field Observations (Dartmouth)
I see a tank!
Query/ Response
Observation Viewer
(9-month demo - standalone)
Registration/ Update Stream
Observations
MBP
(18-month demo - integrated)
6
Field Observations (Dartmouth)
29-SEP-2012 1347.56 OBSERVATION 0018 VEHICLE
Observer 16.35 N, 35.28 E, Elevation 530 m
Sightline 270 deg, 0 deg down, 2000 m Vehicle
Gao, flatbed truck, 3 axles, heading 180,
speed 60 km/h Note 12 soldiers in
flatbed
7
The Challenge Robust Agent Coalitions

• Coalitions are open systems
• Dynamic membership, often novel partners
• Agents in open systems will be unreliable
• Intermittent bugs (3 per 1000 lines in the best
  crafted code) as well as the possibility of
  malice
• Infrastructures can be unreliable
• Current failure tolerance approaches are
  insufficient
• Assume closed systems (e.g. mirroring)
• Full rollbacks are unnecessarily inefficient for
  agents

8
The MIT Robustness Service

• Monitors agent health via polling
• Responds to agent failure via intelligent task
  cancellation task re-announcement
• Maintains reliability information (for failure
  avoidance)
• Designed for open systems - makes minimal
  assumptions about agents

9
A Working Grid Service

• Transparently infers commitment structures
• Assumes (some) agents support (some of ) EH API
• Polling (backup existing Grid is-alive? method)
• Task re-announce
• Cancel-task

10
(No Transcript)
11
Benefits Validated Empirically

• Up to 3x speedup and 8x reduced variability vs.
  standard timeout-retry approach
• Benefits increase with task complexity
  (decomposition tree height) and with level of EH
  API support
• http//ccs.mit.edu/klein/papers/ASES-WP-2000-05.ps

12
Michigan Multilevel Coordinator Agent

• Analyses the alternative plan spaces of coalition
  functional teams that plan independently and act
  asynchronously
• Works top-down with plans chosen by teams to
  predict unintended interactions (resource
  contentions friendly fire).
• Identifies candidate resolutions (timing or
  action constraints).
• Notifies process panel of possible plan
  conflicts and computed workarounds.
• Operationalizes/enforces coordination decisions
  selected.
• Given more time, isolates and resolves conflicts
  more precisely and efficiently.
• Allows planning and coordination decisions to be
  postponed until runtime conditions become better
  known.
• Packaged as a Grid-aware component that will be
  proactively executing and will be utilized by the
  AIAI Process Panel.

13
Michigan Coalition Coordination Example
Forces begin at aircraft carrier AC Airforce
sorties to C, E, Q for Total Exclusion
Zone (TEZ) Logistics delivers humanitarian
aid to refugees at F and R ArmyDiv1 occupies X to
prevent Agadez forces from reaching
and inciting refugees at R ArmyDiv2 crosses TEZ
to occupy Y to monitor for Gao crossings
Potential plan conflicts include friendly fire in
TEZ on ArmyDiv2, destruction of roads through E
that ArmyDiv2 might need, and contention for sea
and rail transport among army divisions and
logistics.
14
Coordinated Plans

• Hierarchical plan coordination incrementally
  recommends coordinated plans that are
  increasingly detailed and parallelized

15
Michigan Multilevel Coordinator Agent