MokSAF: Agentbased Team Assistance for Time Critical Tasks

1
MokSAF Agent-based Team Assistance for Time
Critical Tasks

• Katia Sycara
• The Robotics Institute
• email katia_at_cs.cmu.edu
• www.cs.cmu.edu/softagents

2
MokSAF

• Multi-Agent system that assists human teams in
  time-critical military planning tasks.
• Team goal can be decomposed to two tasks per team
  member
• Coordinated resource allocation
• Geo-spatial planning.
• Two types of agents
• MokSAF Interface agents display individual and
  shared routes, and facilitate communication
  between team members.
• Route Planning Agents (RPA) generate and/or
  critique routes for heterogeneous military
  platoons through a given landscape.
• Research Issues - to investigate different
  approaches for Human/Agent interaction when
  assisting with team goals.

3
MokSAF Team Goal

• Commanders coordinate individual platoons on a
  joint task.
• Each platoon starts in different location, but
  arrives at common rendezvous.

• Have to negotiate
• Composition of each platoon.
• Route taken by each platoon.
• Resource contention (e.g. Fuel).
• Commanders can
• Share details of platoon composition.
• Share route details.
• Re-negotiate rendezvous location and time.
• Discuss strategy through typed messages.

4
Agents within the MokSAF environment

• Consists of three different Task Agents which
  perform route planning.
• Naïve RPA critiques user defined routes
• Autonomous RPA fastest route between two points
• Collaborative RPA refines user defined routes
• Users interact via individual MokSAF Interface
  Agents.
• Interface agent used to construct, view and share
  routes.
• Commanders communicate with each other through
  their personal interface agent.

5
MokSAF Interface Agent
Bravos Shared Route
Intangible Constraints
Tools for constructing routes for the Naïve and
Collaborative agents
Information about team-members routes, platoons
Alphas Shared Route
6
Naïve Route Planning Agent

• Input parameters
• A route represented as a sequence of points.
• Output parameters
• An annotated route (as a sequence of points).
• Naïve RPA critiques route
• Checks route validity
• Identifies constraint violations

7
Autonomous Route Planning Agent

• Input parameters
• Start and end points of a route.
• Output parameters
• The fastest route between these points
  (represented as a sequence of points).
• Factors considered when generating route
• The platoon characteristics w.r.t. type of
  terrain.
• The volume of fuel required to successfully
  achieve the goal.

8
Collaborative Route Planning Agent

• Input parameters
• A corridor represented as a sequence of points
  and a width.
• Output parameters
• The fastest route constrained by the corridor
  (represented as a sequence of points).
• Behavior
• User retains (some) control of route - as with
  Naïve RPA
• Generation of (localized) optimal route - as with
  Collaborative RPA

9
Experimental Tasks

• 3 commanders start at different locations on the
  map.
• Each commander plans a route to a single, shared
  rendezvous point that is to be reached by a given
  deadline.
• Each commander may need to
• Go via one or more fuel depots to refuel.
• Avoid constrained regions, but traverse desirable
  areas.
• Coordinate routes, allocation of vehicles and use
  of fuel depots with other commanders.
• Suggest alternatives to the proposed rendezvous
  point or agreed meeting time.
• Coordination occurs via communication and plan
  sharing.

10
Experimental Results

• The results suggest that commanders were able to
  identify faster (and more economic) routes with
  the aided (ie autonomous or collaborative)
  condition, and hence shared routes faster with
  team mates.
• However, commanders complained about lack of
  control over the final route with the autonomous
  RPA
• Intangible constraints used to coerce RPA.
• Desired route obtained iteratively through trial
  error.

11
Experimental Results

• Commanders identified better routes faster with
  the Cooperative or Autonomous RPA, than with
  Naïve RPA
• This resulted in better coordination with team
  mates.
• Users liked feedback control gained from Naïve
  RPA, but found manual generation of routes
  tedious and slow.
• Users complained about lack of control over the
  final route when using Autonomous RPA
• Intangible constraints used to coerce RPA.
• Desired route obtained iteratively through trial
  error.
• Users expressed preference for Cooperative RPA,
  but use of this agent was not optimal for all
  scenarios.