View
4
Download
0
Category
Preview:
Citation preview
Humans-in-the-Loop Target Classification
Mid-Year Review, Department of Aerospace Engineering,
University of Michigan, April 3, 2008.
Presenter Tom Temple
Joint work with Ketan Savla and Emilio Frazzoli
Laboratory for Information and Decision Systems,Massachusetts Institute of Technology
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 1 / 24
Motivation
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 2 / 24
Problem Formulation
Problem Formulation
The surveillance of a region Q isentrusted to n human operatorsand m UAVs.
Targets arrive randomly in Q ata rate λ with uniformdistribution over Q.The UAVs are routed to visittargets and record video.
Support
System
(SS)
UAVs
m
Human
operators
n
Target
video
Target
video
Target
locations
Target
classification
Target locations
Classified targets
A target is classified when ahuman operator has seenenough video of the target todecide whether it is “lethal” or“benign.”
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 3 / 24
Problem Formulation
Objective
Goals
(i) Design joint motion coordination and operator task allocation policies
(ii) Characterize the quality of service as a function of n and m.
Quality of Service
The average time between the arrival of a target and its classification.
Approach
Theoretical lower bounds, efficient policy design.
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 4 / 24
Problem Formulation
Assumptions
The humans watch video at a real time rate.
Operators can stop and resume where they left off.
Another operators can also resume where a different operator left off.
The an operator’s belief states about a target is not interpretable bythe decision support system.
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 5 / 24
Problem Formulation
The Blame Game
Imagine that a customer is waiting for service at a particular instantblames either the UAVs or the humans for having to wait.
Specifically, if all the human operators are busy at that instant theyblame the humans and otherwise they blame the UAVs.
Define Wv,Wh as the expected integrals of blame for the UAVs andhumans, respectively.
Wq = Wv + Wh
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 6 / 24
Theory
The Light Load case
ρh → 0Wh → 0
Wv is simply the travel time
Wq ∝1√m 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 7 / 24
Heavy Load
Heavy Load: Many UAVs
ρh → 1m→∞
The resulting system is a M/G/nqueue
Wq = Wh =λ(λ−2 + σ2s/n
2)
2(1− ρh) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 100.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 8 / 24
Heavy Load
Heavy Load: Few UAVs
ρh → 1m = n
Reduces to the multi-agent DynamicTraveling Repair-person Problem(mDTRP)
Wq ∝λ
n2(1− ρh)20 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 9 / 24
Heavy Load
Heavy Load: general number of UAVs
ρh → 1Let k = m− n, and hold nfixed.
For the UAVs to be blamed, allk free UAVs must to been-route to targets when anoperator becomes free.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Wv =λ
aon2(1− ρh)2 + a1kn(1− ρh) + a2k2Bounded!
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 10 / 24
Validation
Simulations: Heuristics
A UAV receives a reward whenever a classification is made using videothat they have collected.
UAVs always maximize their expected reward rate for their currentand immediate next target.
When a new target appears, the UAVs bid on it.
A UAV is allowed to leave a target whenever doing so gives a higherrate of reward. This allows for the use of offline information collection.
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 11 / 24
Validation
Simulation
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 12 / 24
movie1s.movMedia File (video/quicktime)
Validation
Simulation
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 13 / 24
movie4.movMedia File (video/quicktime)
Validation
Empirical Results
100
101
102
100
101
102
1
1−ρ
WqOrder of Growth of Waiting Time
m − n = 0
m − n = 1
m − n = 3
m − n = 8
m − n = 13
∝1
1−ρ
∝ ( 11−ρ
)2
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 14 / 24
Summary
Summary
“Human-in-the-loop vehicle routing policies for dynamic environments,”IEEE CDC, Dec 2008.
Provable optimality forsimple algorithms
Provide an analyticalfoundation for generalalgorithms
Addressed how“situational awareness”affects systemperformance
W ∼ (1− ρh)−1
W ∼ 1√m
of ρh for ρh ≤ ρvW independent
Inverse QuadraticTransition in m− n
1
0
W ∼ λm2(1−ρv)2
ρh =λs̄n
ρv =λs̄m
1
W ∼ λn2(1−ρh)2
What if the human operators’ decision time is affected by their load factor?Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 15 / 24
Situational Awareness
Situational Awareness
For the moment, consider a single operator,
We include “Situational Awareness” by letting s = s0χ(ρ)
We can arbitrarily scale s0 such such that the minimum of χ is 1,without loss of generality.
We assume χ is convex, bounded and differentiable on [0, 1]
Since ρ is itself a function of s, the steady state values are thesolution to the coupled equations,
s̄ = s̄0χ(ρ)
ρ =1
λs̄
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 16 / 24
Situational Awareness
Multiple Operators
With multiple operators we are free to choose a solution.
The optimum can be expressed as a straightforward convexoptimization problem
min{ρ1,...,ρn}
∑ni=1 ρiχ(ρi)s̄0
s.t.∑n
i=1ρi
s̄0χ(ρi)= λ
0 ≤ ρi ≤ 1, i = 1, . . . , n.
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 17 / 24
Situational Awareness under Heavy Load
Situational Awareness under Heavy Load
The results presented for the heavy load case are based on queuingarguments that require system stability and do not necessarily carry over.
For fixed s̄, stability is equivalent to ρh < 1.
With variable s̄, this is not necessarily sufficient
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 18 / 24
Situational Awareness under Heavy Load
Two Cases for Real-World χ
If dχ(1)dρ ≤ χ(1) Ifdχ(1)
dρ > χ(1)
ρh < 1 is sufficient for stability
previous heavy load results hold
ρh = 1 is not optimallyproductive
there exists some ρmax < 1 suchthat ρh ≤ ρmax is necessary forstability
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
1.2
ρ
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
ρρmax
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 19 / 24
Situational Awareness under Heavy Load
Perceived load, ρ̃
Defining Situational Awareness in terms of average load, assumes thatthe operators know the average load a priori.
A more realistic model would define Situational Awareness in terms ofperceived load, ρ̃i which is estimated by operator i.
The “error” in this estimation is another potential source of instability.
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 20 / 24
Situational Awareness under Heavy Load
Estimation Stability
0 1 0
1
Percieved utilization ρi
Situ
atio
nal
Aw
aren
ess
Fact
or
ρmax
marginally stableequilibrium
unstableequilibrium
slopes 1λs̄0
Equilibrium Conditions for Operator i
stableequilibria
χ(ρ)
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 21 / 24
Situational Awareness under Heavy Load
Decision Support Implications
We need to carefully manage the perceived operator loads if we are toguarantee system stability.
If for operator i, ρ̃i > ρmax, then we need to give him a break, even ifthere are outstanding targets.
Corresponds to previous results with
n′ ← nρmax and,ρ′h ← ρ/ρmax.
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 22 / 24
Summary
Summary
“Efficient routing of multiple vehicles for human-supervised services in adynamic environment,” AIAA GNC, Aug 2008.
Addressed how “situational awareness” affects system performance
Identified queue stability conditions in heavy load
Determined how and when previous results can be used.
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 23 / 24
Future Work
Future Work
Vanishing targets (customer impatience).
Vehicles with finite capacity e.g. fuel, range, memory.
Alternate support system architectures allowing more involvement byhuman operators in the mission tasks.
Tom Temple (LIDS MIT) Humans-in-the-Loop Target Classification April 3, 2008 24 / 24
Problem FormulationTheoryHeavy LoadValidationSummarySituational AwarenessSituational Awareness under Heavy LoadSummaryFuture Work
Recommended