TEACHING NEW TEAMMATES

1
TEACHING NEW TEAMMATES

• Doran Chakraborty
• Sandip Sen
• MCS Department
• University Of Tulsa

2
Motivation

• Knowledge transfer between agents is challenging
• Different knowledge representations
• Require common background knowledge
• Teaching skill of the trainer
• Learning skill of trainee
• Agent Teaching Agent (ATA) framework
• Expert trainer w/o knowledge of the internal
  representation/algorithms of the novice trainee

3
ATA assumptions

• Trainer does not have access to dataset from
  which it developed its knowledge
• Trainer can introspect to generate problem
  instances of varied difficulty level spanning the
  input space
• Trainer can evaluate the trainees knowledge by
  asking it to solve problem instances
• Trainer can supervise trainee by suggesting
  desired solutions to problem instances
• Incremental, closed-loop training process

4
Agent Teaching Agent Framework
5
Domain Pursuit Problem

• Four identical predators coordinate to capture a
  prey by surrounding it in continuous space
• Prey moves linearly at half the speed of the
  predators
• Predators can see others and the prey but cannot
  communicate to coordinate
• Predators in close proximity (dc) collide and
  recoil a significant number of steps away from
  the collision point
• Predators must capture the prey within a fixed
  number of steps

6
Predator strategies

• Experts who have a prior knowledge about the
  strategy
• A predator sensing another predator in close
  proximity (dp) moves orthogonally to the
  direction of the prey.
• Trainees Use Instance Based Learning (IB2)
• to learn concept description from trainers.
• Trainers Use Decision tree learning algorithm
  (C4.5) to learn concept description from experts.

7
Coordination knowledge repository

• Exemplars Agent positionsrecommended action
• Experts provide the knowledge repository.
• Simulating representative scenarios
• Challenge determine a small set of
  scenarios to facilitate training of the entire
  range of desired behavior?
• Our Approach
• All possible placement of predators in four
  quadrans centered on the prey

8
Possible Set 1
9
Possible Set 2
10
Possible Set 3
11
Training set

• ltr1, ?1, r2, ?2, r3, ?3, rprey, ?prey, Mgt
• (r1 , ?i) polar coordinates of predator i wrt
  the expert predator
• (rprey , ?prey) polar coordinates of the prey
  wrt the expert predator
• M move recorded

12
Agents

• E expert
• C trainer
• I an instance based learner (trainee) trained
  in one shot by the trainer
• A an instance based learner (trainee) trained
  by the trainer using the ATA framework.

13
Experimental Set Up

• Game played over 20 20 grid
• Results averaged over 1750 games.
• 7 compositions of predators
• Homogeneous groups
• EEEE all experts
• CCCC all trainers
• IIII all I trainees
• AAAA all A trainees
• Heterogeneous groups (varying number of trainees)
• CCCA
• CCAA
• CAAA

14
Results
15
Summary over 1750 games
16
Highlights

• AAAA perform much better than IIII, and close to
  CCCC
• Varying the trainer trainee composition in the
  team
• Injecting more trainers increases the performance
  level of the group
• Interestingly heterogeneous groups performing
  better than homogeneous groups Can be due to the
  complementarity in the knowledge of C and A.

17
Future Work

• Use more knowledge representations, e.g., SVM
• Colluding predators
• Challenging knowledge transfer scenarios
• Robocup soccer
• Coordinated surveillance/monitoring/scouting