samad/confs/artist02/ high-confidence control ensuring reliability in high-performance real- time...
TRANSCRIPT
Samad/confs/Artist02/
High-Confidence Control
Ensuring Reliability in High-Performance Real-Time Systems
Tariq SamadHoneywell Automation and Control Solutions
Minneapolis, U.S.A.
Collaborators: Pam Binns, Mike Elgersma, Vu Ha
Research supported by DARPA/AFRL contract F33615-01-C-1848
Samad/confs/Artist02/
Increasing Autonomy in UAVs (USAF Perspective)Increasing Autonomy in UAVs (USAF Perspective)
Samad/confs/Artist02/
Limitations of “Automation”Limitations of “Automation”
• Human operators are not needed for nominal conditions today
– straight-and-level cruise flight in good weather
– steady-state operation of process plants
– urban roadways under light traffic flows
• Invariably, abnormal conditions require human intervention
• Autonomy implies appropriate responses to unforeseen situations
– all control behaviors cannot be pre-compiled
• Much research in systems and control is focused on enabling autonomy
– but there’s a theory/practice gap
Samad/confs/Artist02/
The Theory/Practice Gap in ControlThe Theory/Practice Gap in Control
• Several new theoretical and analytical developments in systems and control over the last decade or two
– nonlinear control
– intelligent control
– adaptive control
• Notable successes in practical applications, but the full potential of these and other techniques hasn’t been realized
• The problem isn’t incomplete theory, or a lack of simulation results
• The problem is the lack of “confidence” for real-time, life- and mission-critical applications
• Human operators are not primarily employed for performance, but for confidence
Samad/confs/Artist02/
Difficulties with DeterminismDifficulties with Determinism
• Current verification and validation (V&V) approaches are infeasible for future systems
– focus on deterministic guarantees of safety
– complex algorithms are analytically intractable
– exhaustive analyses are impossible
• Promising alternative: probabilistic methods
– algorithmic performance measures (inc. control stability)
– Reliability and dependability analyses
– Probabilistic online admission control
– Statistical verification of execution properties: focus of our work
Samad/confs/Artist02/
The Verification Problem for Advanced ControlThe Verification Problem for Advanced Control
• For safety- and mission-critical systems, verification practices today focus on exhaustive, worst-case analyses
– e.g., ensure, under all conceivable conditions, that the calculation will complete within the deadline
• Computationally sophisticated algorithms are either avoided entirely, or only used in restricted, provably safe situations
– nondeterministic execution times
– computations performed depend on state and inputs
• Real-time control applications impose hard deadlines on computation
– difficult to guarantee that calculation will be completed by deadline
Samad/confs/Artist02/
Performance with ConfidencePerformance with Confidence
The conservative, deterministically verified region of acceptance:
xmin < x < xmax
Unacceptable Acceptable
Continuous dimension x1
Discretedimension x2
Characterization of computational assurance
The new,statistically verified region of acceptance
Unacceptable AcceptableDiscrete
Characterization of computational assurance
Continuous dimension x1
dimension x2
0)(0)(0)()( 321 xdxdxdxh
Example problem: computation of trim solution
Iterative computation for x may or may not converge within deadline
Convergence a function of state and inputs
),(const uxfxtrim
Samad/confs/Artist02/
Asymmetric PenaltiesAsymmetric Penalties
• In most cases of interest, some degree of conservatism will likely be desirable
h1
h3
h2
XHypothesis h1 tolerates false positives in the
interests of high performance
Hypothesis h2 is “minimally safe” for the given data set
Hypothesis h3 results in conservative decisions
Samad/confs/Artist02/
Some Basic SLT ResultsSome Basic SLT Results
• For any probability distribution D on X {-1,0}, any hypothesis h in H that makes k errors on a training set of m random examples will have a generalization error probability bounded as follows (the result applies with probability 1- and assumes that d m):
4
ln2
ln42
d
emd
mm
kR
• Assume we observe an empirical error (more generally, risk) of classification, Remp, based on
a “training set” of m examples for classifiers from a hypothesis space H. The statistical learning theory model formulates how Remp differs from R:
1)()(supPr hRhR emp
Hh
• The result above is assured under fairly general conditions provided that (where d is the VC dimension of the hypothesis space H of classifiers we are considering):
13
log82
log41
22 dm ~Independent of problem dimension!Assumes “consistency” of H
Specialized for classification problems; consistency not assumed
A distribution-free formulation!
Samad/confs/Artist02/
VC DimensionVC Dimension
• Intuitively, a measure of the flexibility or richness of a hypothesis space
• Examples:
– lines in R2 have VC-dimension of 3
– half-planes in Rn have VC-dimension n+1
– axis-aligned hyperrectangles in Rn have VC-dimension 2n
– n-sided polygons in R2: 2n+1
– k-sided convex polyhedra in Rn: 2klog2(ek)(n+1)
Samad/confs/Artist02/
A Methodology for Statistical VerificationA Methodology for Statistical Verification
• Given a data set of m samples {(x1, y1), (x2, y2), …, (xm, ym)}
• Find hypothesis in hypothesis space H of VC dimension d with
– zero false positives (i.e.,wrong prediction of computing feasibility)
– low false negatives (I.e., wrong prediction of computing infeasibility)
• Calculate ub as
ub is an upper bound (with confidence 1-) of the true probability
for a false positive
• Given H and confidence level , we can estimate safety as a function of m and vice versa
• Allows explicit tradeoff between safety and performance
m
dmdub
4/ln1/2ln
Samad/confs/Artist02/
OAV Features and FunctionsOAV Features and Functions
Navigation performance:
• GPS has 20 ft. error: landing area must be 40 ft. in diameter and flat.
• Station location keeping is within 5 to 10 ft.
• Altitude is known within 5 ft.
Flight
• AV2 endurance: 95 minutes
• Speed: 100kts
• Auto start, take-off/landing
• Waypoint designation at any resolution (direction, distance, speed).
• Blind descent: 1ft/sec. descent with weight-on-wheel sensors; terrain data not required.
• Flight control allows for hover and translation maneuvers.
Samad/confs/Artist02/
OAV ApplicationOAV Application
• OAV has lift surfaces in propulsion airflow
– causes significant nonlinear interactions between thrust and surface variables
• Requires the online calculation of equilibrium angle of attack
• Initial dependency explored: net lift force and flight path angle
– net lift force = 2S/2mg (: air density, : speed, S: effective wing area)
– flight path angle () is zero for upright flight
• Currently working on 4-D problem
– includes body-axis unit vector
Samad/confs/Artist02/
Application in ProgressApplication in Progress
Upper bound on probability of unsafe condition is 0.045, with 95% confidence
Can increase safety with more samples or lower VC dimension m = 700000
VC dimension = 137 = 0.05 (95% confidence)ub = 0.045
• Iterative equilibrium angle of attack computation for new small-scale UAV• 2-d slice of computational complexity shown, as a function of two inputs to the computation• Characterize region of performance where 2 iterations suffice (red region in graphs)• Hypothesis space: logical combination of four
quadratics• Realistic 4-d application in progress
Samad/confs/Artist02/
ConclusionsConclusions
• Automation is pervasive, autonomy is not
• High-performance algorithms aren’t sufficient; higher confidence implementations are needed
• One problem/solution: Statistical verification of advanced control software
• Many other possibilities to close the theory/practice gap
• Exciting opportunities for both research and impact!
• Multidisciplinary efforts essential