Using Games and Agents in C2 Research InProgress Briefing

1
Using Games and Agents in C2 Research
In-Progress Briefing
Linking People and Technology

• Marcy Stahl
• mstahl_at_thoughtlink.com
• 703/820-8227
• June 19, 2002

2
Background

• Sponsor Joint C4ISR Decision Support Center
  (DSC)
• PIs ThoughtLink and Center for Naval Analyses
  (CNA)
• DSCs task Develop a Joint Force Command and
  Control operational concept, operational
  architecture, and implementation roadmap to
  achieve the goals of Joint Vision 2020
• TLI/CNA Objective Use game-playing agents and
  humans to explore C2 design characteristics
• TLI Lead on human games and C2 literature review
• CNA Lead on agent-based modeling

3
Why Games?

• Well-designed games are powerful abstractions
• Reduce complexity
• High fidelity does not imply high validity
• Enable statistical analysis
• Theyre fun!
• Helps engage human participants

4
(No Transcript)
5
The SCUDHunt Game

• Game provides an operational back story for
  context
• 3 SCUD launchers hidden in the hostile
  (hypothetical) state of Korona
• Game played online by 4-person distributed teams
• Each player controls 1-2 information assets
  (satellite, Navy seals, etc.)

• On-line, instrumented version
• Captures info for assessing distributed teams
• Team members must collaborate
• Plan their search, strike plans
• See http//www.scudhunt.com for more info

Sample game board
6
Key Measures
Shared Situational Awareness (SSA) - overlap in
assessment of launcher locations among team
members, irrespective of whether understanding is
right or wrong SSA Ratio of the total number
of recommended target squares by all players to
total number of unique squares designated Example
Perfect SSA All 4 team members vote for the
same 3 squares 12/3 4 Lowest score All 4
team members vote for 3 different squares 12/12
1
7
Key Measures - Cont.
Accuracy (ACC) - Do team members (or individual
players) actually find the launchers? ACC
ratio of nominated squares that actually
contained SCUD launchers to the total number
squares nominated Example Perfect team ACC 4
players vote for the same 3 squares containing
launchers 12/12 1. Lowest ACC Team does not
identify any launcher squares, then their score
is 0 / 12 (or some other large number) 0 We
also compute individual player ACC
8
Other SCUDHunt ExperimentsExploring SSA

• TLI/CNA experiment for DARPA, 2000
• Effect of visualization and communication modes
• Additional data mining sponsored by DSC, 2001
• University of Arizona, 2001
• Effect of leadership and asset reliability
• Naval War College (NWC), 2002
• Effect of command method and visualization mode
• Supported by TLI and CNA
• Army Research Institute and George Mason
  University, 2002
• Effect of player cross-training

9
Returning to our regular programming...

• What are we and CNA doing for DSC in 2002?

10
Researching Inter-relationships among C2 Design
Characteristics

• 1. Timely Preparation and Training
• 2. Broad Scope
• 3. Shared Understanding
• 4. Enterprise-wide Integration
• 5. Comprehensive Decision Support
• 6. Superior Decision-Making
• 7. Flexible Self-Synchronization
• 8. Flexible Dispersal
• 9. Simultaneous C2 Processes
• 10. Simultaneous and Continuous Operations
• 11. Agility

11
SCUDHunt and the real world
12
Human and Agent games
Explore actions and rationales Generate many
cases
Agent Game
Interesting patterns
How human players act
Do humans act that way?
Human Game
13
Agent Simulation

• CNA Team Dr. Peter Perla, Dr. Andy Ilachinski,
  Dr. Carol Hawk, Michael Markowitz, Chris Weuve
• Approach
• Use agents as substitutes for humans to explore
  large design space
• 4 agents collaborate to play SCUDHunt as a team
• Agents have different personalities
• Compute SSA and ACC metrics
• Identify interesting areas for addl human games
• Future work use genetic algorithms to understand
  fitness functions in real-world C2 problems

14
Agents and their jobs
Agent actions
Agent info
Place assets
Interpret search results
Determine what info to share with other agents
Update beliefs
Submit strike plan
15
Agent basics

• State of the game
• Belief-matrix, -1 ? Bijt ?1 how strongly agent
  believes SCUD launcher is in grid i,j at turn t
• Agents beliefs change in response to
• Agents interpretation of search results
• Agents perceived reliability of asset
• Other agents communicated beliefs
• Agents trust in other agents, and

16
Agent basics - cont.

• Agents beliefs change in response to
• Threshold for strike plan nominations
• Threshold for placing assets
• Fitness function with weights to determine asset
  placement
• Agents beliefs and thresholds represented by
  matrices
• Belief matrix updated by multiplying various
  matrices together and then doing a Durkin fuzzy
  sum with current and prior beliefs

17
Agent Outputs to Consider

• SSA scores
• Team ACC and individual agent ACC scores
• Belief matrix values

Sample belief matrix from initial team of agents
18
Carbon-based SCUDHunt Experiment

• Goal Explore effect of visualization and quality
  of information on SSA and ACC
• Experiment design
• Latin square with 6 teams playing 6 games in
  various sequences
• 2 levels of visualization (shared viz of actual
  sensor data and post viz of interpreted sensor
  data)
• 3 levels of quality of information (LOW false
  positives and false negatives MEDIUM false
  negatives but no false positives HIGH small
  prob. of false negatives)
• Schedule July-August 2002 at Naval War College

19
C2 Campaign Plan Increase Complexity as Problem
Comprehension Increases
Gaming Environments
Subjects
CPX
SCUDHunt-A
LOEs
SCUDHunt
Adaptive Agents
Game Complexity
Feedback Loop
Students/ General Population
Access/Scheduling Difficulty
Realism
Variables

• Tools (comms, viz)
• Team factors
• Size
• Heterogeneity
• Personality types
• Processes
• Addl C2 Design Characteristics

Feedback Loop
Real-world Interagency Personnel
SCUDHunt Adapted
20
Summary

• Abstract games have proven very useful for C2
  research
• Initial agent actions are promising
• Our agent and SCUDHunt experience will be a
  foundation for developing a campaign plan for
  broader C2 research agenda
• For more info on topics discussed, see
  http//www.scudhunt.com or http//www.thoughtlink.
  com

21
Agents and sensors

• Interpretation of sensor reports
• Sensor-ReportLauncher-Correlation Matrix
• bRS Agents belief that launcher is at
  coordinate for which sensor S has
  reported R
• Sensor Reliability Estimate Matrix
• RRS As estimate of the reliability of
  sensor Ss report R
• 0 ? RRS ? 1

22
Agents and other agents

• Trust (of other agents)
• Agent?Agent Trust Matrix
• 0 ? TAB ? 1
• TAB 0 agent A mistrusts everything agent B
  tells it
• TAB 1 agent A believes everything agent B
  tells it
• Belief Update
• Own Sensors BownRRS bRS
• Linked Sensors Blinked TAB RRS bRS or
  Blinked TAB BL,ij,
• where BL,ij is the belief matrix of agents linked
  to A

23
Agent beliefs

• Belief Update (using Durkin fuzzy-sum)
• Bij(t1) Bij(t)?Bown(t) ? Blinked(t), where

24
Strike plans

• Strike Plan Logic
• Select top NStrike ranking sites
• such that Bij Bthreshold
• where 0 ? Bthreshold ? 1 is As Threshold
  Belief Strength
• Bthreshold ? 0 ? A is easily convinced
• Bthreshold ? 1 ? A is stubborn

25
Sensor placement

• Sensor Placement Logic
• Dogma Threshold, 0 BDogma 1
• If Bij BDogma then A places a launcher is
  definitely here marker at site (i,j)
• If Bij - BDogma then A places a launcher is
  definitely not here marker at site (i,j)
• Sensor Placement Fitness Function

26
Variables to explore