CAST:%20Collaborative%20Agents%20for%20Simulating%20Teamwork

1
CAST Collaborative Agentsfor Simulating
Teamwork
John Yen, Jianwen Yin, Thomas R. Ioerger,
Michael Miller, Dianxiang Xu, Richard
Volz Department of Computer Science Texas AM
University
Acknowledgement DoD/AFOSR MURI grant
2
Our Goal Building IntelligentTeam-Training
Systems

• Not just better individual performance, but
  better coordination, load-balancing, situation
  awareness
• Examples fire fighters, air traffic controllers,
  TOCs, CICs, AWACS, shuttle mission control
• Scenario-based training via distributed
  simulation
• The need for agents - automated coaches and role
  players (virtual team members)
• Agents must understand the team structure and
  teamwork process

3
Studies on Team Trainingfrom Psychology
Literature

• teamwork vs. taskwork (Salas Cannon-Bowers)
• process vs. outcome measures
• communications frequency, type (Serfaty)
• situation awareness (Endsley)
• adaptiveness of team to workload
• (also personality, leadership...)
• training protocols like cross-training to learn
  other roles (Salas Cannon-Bowers)

4
Roles for Agents in Team-Training

• Role players/Virtual Team members
• reduce cost of training (e.g. need for human role
  players)
• need to know when to act and when to defer
• need to know whom to coordinate with
• need to know whom to share info with
• must describe team task (plan) and role
  assignments
• Automated Coaches
• build student model (observe actions infer their
  view)
• compare to expert model (what should the
  trainee do ideally, given the team structure?)
• give feedback, critique, instruction, more
  training...

5
Key Concept Shared Mental Model

• Various components
• static structure of the team, comm. policies...
• goals and plans
• dynamic current situation, others
  workloads/status
• Needed for virtual team members
• not just domain knowledge
• also roles, responsibilities, capabilities, team
  plans
• need to know who should act and when
• need to know when to communicate for sync.
  coordination, disambiguation, infomation sharing,
  etc.

6
Shared Mental Models, continued

• Needed for user modeling
• how to interpret incorrect actions?
• alternative explanations of observed failures of
  action
• thought it was anothers responsibility?
• waiting for synchronization or approval?
• meaning of shared responsibilities delegation
  for backup behavior (important for robustness)

7
Related Work

• STEAM (Tambe)
• STEVE (Rickel Johnson)
• REACT (Hill Johnson)
• SWARMM/dMARS (Tidhar Jennings)
• lots of others...

8
The CAST Agent Architecture

• MALLET - team KR language
• team structure
• team process
• CAST kernel (interpreter)
• convert to Petri nets (track progress, select
  actions)
• use back-chaining theorem-prover for inference
• dynamic role selection - make choices in context
• DIARG - information exchange algorithm
• proactive offer new info to those who need it

9
MALLET Multi-Agent Logical Language for Encoding
Teamwork

• syntax based on S-expressions (symbolic, nested
  lists)
• basic definitions
• (team ltteamNamegt (ltagentNamegt))
• (role ltroleNamegt)
• (plays-role ltagentNamegt ltroleNamegt)
• (capable ltagentNamegt ltoperNamegt)
• conditions (ltpredicategt) with variables
  prefixed by ?
• e.g. ((forward-scout ?unit) (location ?unit ?x
  ?y))
• operators
• (indiv-oper ltoperNamegt (ltvargt)
• (pre-cond ltcondgt) (effects ltcondgt))
• (team-oper ltoperNamegt (ltvargt)
• (share-type ANDORXOR)
• (pre-cond ltcondgt) (effects ltcondgt))

10

• team plans
• (team-plan ltplanNamegt (ltvargt)
• (pre-cond ltcondgt) (effects ltcondgt)
• (term-cond ltcondgt SUCCESSFAILURE)
• (role-select ltvargt (ltroleNamegt)
• (constraint ltcondgt))
• (process ltprocessgt))
• process expressions
• ltprocessgt (seq ltcallgt) (par ltcallgt)
• (while ltcondgt do ltprocessgt)
• (if ltcondgt then ltprocessgt else ltprocessgt)
• (do ltagentNamegtltroleVargt ltcallgt)
• where ltcallgt (ltplanNamegtltoperNamegt ltarggt)
• semantics of responsibilities - similar to joint
  intentions (mutual belief) but asymmetric
  (Ioerger and Johnson, IC-AI 2001)

11

• (team-plan T1 ()
• (process (par (kill-wumpuses) (collect-gold))))
• (team-plan kill-wumpuses ()
• (role ?s (scout)) (role ?f (fighter))
• (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 ()
• (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
12
CAST Kernel

• compile team plans into Petri nets (incl.
  expanding sub-tasks)
• cycle sense/decide/act loop
• 1. update beliefs about environment in selfs KB
• 2. check for any incoming messages from other
  agents
• 3. find active steps in plan (transitions with
  tokens in all input places)
• 4. if self is uniquely resp., consider executing
  oper.
• 5. if oper is XOR and resp. is ambiguous, offer
• 6. if oper is AND, broadcast READY and wait for
  others
• 7. randomly choose among remaining actions and
  execute
• 8. inform others of completed steps
• Dynamic Role Selection (DRS)
• check role definitions, must satisfy any
  constraints, capable?
• communicate when ambiguity exists
• sync. for AND operators select for XOR operators
• could also allow individuals to vote/negotiate

13
DIARGDynamic Inter-Agent Rule Generator

• Info. sharing is a key to flexible teamwork
• more generally distributed SA
• training target learning what is relevant to
  whom?
• Want to capture information flow in team,
  including proactive distribution of information
• Want to restrict to only the most relevant cases
  (Tambe)
• Ideal criteria
• (Bel A I) (Bel A ?(Bel B I)) (Bel A (Goal B
  G)
• ?(Bel B I) ? ?(Done B G)
• (Bel B I) ? ? ?(Done B G)
• ? (Goal A (Inform B I))
• where is the temporal operator for always

14
DIARG, continued

• Explanation - A should send message I to B iff
• A believes I is true
• A believes B does not believe I (or believes it
  is false)
• I is relevant to one of Bs goals
• i.e. pre-cond of current action that B is resp.
  for in team plan,
• and that action would not succeed without knowing
  the info.
• Algorithm
• 1. check for transitions which other agents are
  responsible for that can fire (pre-conds
  satisfied)
• 2. infer whether other agent might not believe
  pre-conds are true (currently, beliefs based on
  post-conditions of executed steps, i.e. tokens in
  output places)
• 3. send proactive message with information

15
Experimental Testbed

• Wumpus World (Russell and Norvig) extended to
  teamwork environment
• several agents exploring a 10x10 cave
• goals collect gold, kill wumpuses
• assume variable wumpuses, but stationary
• Roles
• scouts/climbers - look for (smell for) wumpuses,
  can climb over pits
• shooters/fighters - have bow to shoot wumpus
  (from adjacent room), must collect arrows
• carriers - strong for carrying gold

16
Experiment 1

• Evaluate effect of teamwork and information
  exchange on team performance
• Team A CAST, using DRS and DIARG
• Team B CAST - no DIARG, just broadcast all new
  info.
• Team C no teamwork - agents wander randomly

138
17
Experiment 2

• Effect of Dynamic Role Selection
• constraint choose closest fighter to wumpus
• test scalability via
• increasing of pits (in 10x10 cave)
• makes it harder for agents to navigate

18
Future Work

• Add delegation (role/resp change over time)
• contract about communication between A and B
  (maintain mutual belief about goal status)
• Info. exchange (DIARG) depends on frequency of
  change and observability
• More complex model of agent capabilities
• degrees of success (performance scores)
• depends on workload, individuals skill,
  deadlines, dual-tasks
• affects team decision-making

19
Future Work, continued

• Dynamic planning
• generate plan if pre-conditions of current action
  are unsatisfied (instead of just waiting)
• how will this impact team modeling?
• Interaction with humans
• must infer their view of the teams progress
• deception, untruthfulness
• failures due to forgetting/overload
• conflicts with private goals?

20
Conclusion

• Introduced CAST agent architecture
• MALLET team-representation language
• Major algorithms
• dynamic role selection (in kernel)
• proactive information exchange (in DIARG)
• Simulation of a shared mental model
• Working toward support of intelligent
  team-training systems (user-modeling coaches, and
  virtual team members)