Modeling Capabilities and Workload in

1
Modeling Capabilities and Workload in
Intelligent Agents for Simulating Teamwork
Thomas R. Ioerger, Linli He, Deborah Lord Dept.
of Computer Science, Texas AM University Pamela
Tsang Dept. of Psychology, Wright State University
2
Teamwork and Team Training

• Examples fire fighters, air traffic controllers,
  military, sports, businesses
• Characteristics of effective teams
• commitment to shared goals
• communication, information sharing
• coordination, synergy, non-interference
• back each other up in case of failure
  (robustness)
• flexible, adaptive

3

• Training methods
• practice learn team structure (roles) and
  process (procedures, policies)
• build shared mental model cross-training
• stress innoculation (adaptive, load-balancing)
• Intelligent Agents can help!
• distributed simulations (with cognitive fidelity)
• act as role players (teammates or others)
• can monitor/evaluate teamwork and act as coach by
  giving feedback (dynamic or AAR)
• Our goal is to develop psychologically-sound
  methods for using agents to train teams

4

• The importance of reasoning about capabilities
  and workload in teams
• load-balancing, adapting to shifting task demands
• self-assessment when to ask for help?
• who to ask for help?
• requires knowledge of capability of others
• also depends on their workload (best available)
• awareness of load on others
• when not to interfere, disrupt suppress
  communication
• e.g. air traffic controllers
• can also offer to help (proactiveness)
• How to get agents to understand this?

5
Toward A Computational Model

• Motivating observations
• humans can perform multiple tasks in parallel
• humans have internal limits on processing
  capacity
• humans can often get better performance by
  applying more effort (within limits)
• humans can complete tasks faster by more effort
• Therefore...
• a human is capable of doing a task within a
  deadline if there is enough reserve capacity for
  the effort required
• definition of capability for set of tasks
  depends on finding a schedule such that overlaps
  do not exceed capacity
• options are to delay processing, or stretch out
  a task over longer time to make more resources
  available for other tasks

6

• We assume there is a single resource (similar to
  attention)
• We assume it is bounded umax
• umax may differ between individuals, or with
  fatigue, etc.
• efforttotal resource required over
  timee(T)?t1..t2 et(T)
• let e be average momentary effort e(T)(t2-t1).e
• in some cases, greater effort leads to better
  performance
• Performance Resource Functions
• resource-limited tasks
• workload is sum of effort being applied to all
  tasks at a given moment wtSi et(Ti)
• must satisfy capacity constraint at all times
  umax?wt

Umax
effort iso-curve
average resource utilization
resource utilization
e
task 1
time
duration
qF(e)
quality or performance
effort (res. util.)
7

• A schedule is an assignment of a start time, end
  time, and average effort level for each task in a
  set lttstart,tend,egti
• Main Definition an entity is said to be capable
  of doing a set of tasks T1...Tn if there exists a
  schedule lttstart,tend,egti such that
• 1) meet each deadline dl(Ti) ? tend(Ti)
• 2) enough effort for required quality
• F(ei.(tend-tstart)) ?q(Ti)
• 3) never exceed capacity ?t umax ? wtSi et(Ti)

start end
Umax
resource utilization
T1 T3 T2
e
T4
time
8
Comments

• Interruptibility of current tasks?
• Computational Complexity
• scheduling is NP-hard
• exponential in tasks to solve exactly
• in practice, tasks is small
• do humans use heuristics, like longest-first?
• Extension to Multiple Resources (ala Wickens)
• multiple resource pools, each with own limit
• tasks use profile of resources scales with
  effort
• explains differential interference by task types

9
An Agent Could Use this Model of Capabilities...

• to monitor/assess a subjects workload
• to explain observed performance decrements
• to select training events to push a students
  abilities
• to interact appropriately as a teammate
• offer to help those who need it
• minimize communication to avoid distraction
• to evaluate team performance (load-balancing)
• to augment the shared mental model of team

10
Weaknesses and Limitations

• Integration with discrete tasks?
• How valid are additivity and tradeoff
  assumptions?
• Managing priorities which task to drop, if no
  feasible schedule exists?
• How to determine parameters?
• empirically, e.g. via secondary task interference
• evaluation method (work-in-progress)