Agent-based Proactive Teamwork

1
Agent-based Proactive Teamwork

• John Yen
• University Professor of IST
• School of Information Sciences and Technology
• The Pennsylvania State University

2
Motivation

• Psychological Studies about Effective Human
  Teamwork Indicated that
• Team members can anticipate needs of team mates
• Team members can offer relevant information
  proactively.
• These teamwork behaviors are based on an
  overlapping shared mental model.

3
A Response to the Challenges

• Empower agents with shared mental models" to
  enable them to better
• Anticipate information needs of teammates in a
  dynamic environment
• Offer relevant information proactively to
  soldiers

4
Shared Mental Model

• Shared Ontology
• Shared Goals
• Shared Team Structure
• Shared Team Collaboration Process
• Shared Belief about the Team
• Shared Belief about the World
• Shared Hypotheses about the Enemy

5
Major Features of CAST

• Use a high-level language to describe teamwork
  knowledge
• Capture shared mental model about team
  structure and process
• Infers information needs (from SMM)
• induces proactive information exchanges

6
Anticipating Information Needs of Teammates
Dynamic Task Allocation
Team Plan
Responsibilities of Tasks
Preconditions of Tasks
Who
needs what
Information Needs
7
Dynamic Task Allocation
Team Plan
Roles of Agents in the Team
Dynamic Task Allocation
Constraints for Task Allocation
My Belief about The World
My Belief about Teammates
8
Proactive Information Delivery
Information Needs
Information
Match ?
My Belief about Teammates
Does he/she know ?
How to inform him/her?
Communication Strategy
9
CAST Agent Architecture
Belief Update
Information
Team Knowledge (MALLET)
Responsibility Selection
Belief
Responsibilities (Petri Nets)
Domain Knowledge
Identify Info Needs
Information Needs
Act on Info Needs
10
Shared Mental Model in CAST

• Prolog knowledge base belief
• MALLET High-level language for representing team
  knowledge
• Petri Nets An agents internal representation of
  the dynamic teamwork processes and related
  information requirements

11
Relationships between SMM Components
MALLET Compiler
MALLET Knowledge Base
Petri Net (team process)
query
Prolog Knowledge Base (belief)
CAST Kernel
reply
12

• (team-plan T1 ()
• (process (par (kill-wumpuses) (collect-gold))))
• (team-plan kill-wumpuses ()
• (agent-bind ?s (play-role ?s scout))
• (agent-bind ?f (play-role ?f fighter)(closest
  ?f wumpus))
• (process (while ((wumpus ?x) (not (dead ?x))))
• (seq (do ?s (find-wumpus ?x))
• (do ?f (move-to-wumpus ?x))
• (do ?f (shoot-wumpus ?x)))))
• (team-plan find-gold ()
• (agent-bind ?c (play-role ?c carrier))
• (process (while (true) (if (see ?any-agent
  glitter)
• (do ?c (carrier-pickup gold))))

wumpus exists
start
find
shoot
move
done
pickup
glitter
no wumpuses left
13
CAST Software Architecture
Team Process Execution
Team Process Specification
Monitor
Goal Selection
MALLET
Process Net
Kernel
Knowledge Base
Communication
Parser
Inference Engine
Team Structure
Agent Selection
Domain Adapter
Domain Simulator
Domain World
Inference Engine
Fact Base Domain Policy
Monitor
14
CAST Development Environ.
Circles are places and hold tokens denoting
current execution state. Red indicates the
presence of a token. Rectangles are transitions
and are tested and executed when preceding places
have tokens.
15
Theoretical Foundations of CAST

• What are the principles that derive information
  needs?
• What is the semantics of proactive information
  delivery?
• How does an agent choose among different
  proactive communication strategies?

16
Agent Theories

• Joint Intention theory (Cohen Levesque)
• SharedPlan Theory (Grosz Kraus)

17
Two Types of Information needs

• Action-performing information needs
• enables an agent to perform certain (complex)
  actions, which contributes to an agent's
  individual commitments to the whole team.
• Goal-protection information needs
• allows an agent to protect a goal from potential
  threats that may result in a conflict with the
  goal.

18
Semantics of Assert Performative (Cohen
Levesque)

• The speaker attempts to establish a joint belief
  with the receiver that
• The speaker believes the information
• Whether the receives accept the belief depends on
  the reply (accept or reject)

19
Semantics of ProAssert

• The speaker attempts to establish a joint belief
  with the receiver that
• The speaker believes the information
• The speaker believes that the receiver needs the
  information

20
Semantics of ProAssert

• An attempt for the speaker (s) to establish a
  joint belief that
• S believes in the information (p)
• S believes that the receiver (a) needs the
  information.

21
Responses to ProAssert

• Three possible responses to ProAssert
• Accept The receiver believes the information
• Reject1 The receiver refuses the information
  because the information is contradictory to its
  beliefs.
• ?Reject2 The receiver refuses the information
  because it believes it does not need the
  information

22
Inferring Proactive Assert from Information Needs

• Derived from the Assist Axiom in SharedPlan
  theory

23
Choosing Communication Policies

• How to inform teammates? Proactive or wait for
  them to ask?
• Factors that affect the policies
• Cost of communications
• Certainty level about teammates needs
• Consequence of communications
• Criticality of the information to teammates
• Load level of teammates
• These factors depend on situations agents are in

24
Decision-theoretic Approach to Communication

• Use decision theory for choosing communication
  strategies
• Assign utility to plans
• Assess the value of information from the utility
  of plans
• Specify probabilities in JARE

25
Experiment

• Goal Evaluate Two Proactive Communication
  Strategies

Strategy Action-performing Information Needs Goal-protection Information Needs
S1 Always ProInform Always ProInform
S2 Decision-theoretic ProInform Decision-theoretic ProInform
26
MALLET Example

• (action attack-enemy ()
• (pre-cond (is-enemy_homebase ?h1)
• (at ?h1 ?x ?y) )
• (process
• (par
• (do scoutteam seeking)
• (do fighterteam attacking-wumpus)
• (do boomerteam attacking-base ?x ?y)
• ) ) )

27
Petri Net Process Representation

• Parsed from MALLET
• Hierarchical structure
• Extended with predicate
• Support basic processes such as sequential,
  parallel, and conditional

28
Domain Overview

• Two teams, attack team (blue) and defense team
  (red)
• The mission of the blue team to is to destroy
  enemy base by flowing certain waypoints
• Communication risk
• Different observability
• The red team will patrol around the waypoints and
  home base

29
Domain Simulation Interface

• Blue team
• Red team
• Waypoint
• Enemy Base
• Agent Monitor
• World Monitor

30
Demo
31
Experiment Result
32
Applications

• Training for AWACS-like synthetic task (AFOSR
  MURI)
• Support Horizontal Fusion STO (Army Research Lab)
• Negotiation among Agent Teams for Design (NSF)
• Supporting Analyst Teams (Lockheed Martin)

33
Conclusion

• A computational shared mental model is critical
  for developing agents that support proactive
  teamwork.
• CAST enables proactive information delivery by
  anticipating needs of teammates.
• MALLET facilitates the reuse of teamwork
  knowledge
• CAST achieves efficiency using shared team plans
  and shared policy
• CAST achieves adaptability by dynamic assignment
  of agent responsibility