LQR Performance Index Distribution with
Uncertain Boundary Conditions
Marcus J. Holzinger
Graduate Research Assistant
Aerospace Engineering Sciences
University of Colorado at Boulder
Daniel J. Scheeres
Professor, Seebass Chair
Aerospace Engineering Sciences
University of Colorado at Boulder
Abstract—The track assignment problem in applications
with large gaps in tracking measurements and uncertain
boundary conditions is addressed as a Two Point Boundary
Value Problem (TPBVP) using Hamiltonian formalisms.
An L2-norm analog Linear Quadratic Regulator (LQR)
performance function metric is used to measure the trajec-
tory cost, which may be interpreted as a control distance
metric. Distributions of the performance function are
determined by linearizing about the deterministic optimal
nonlinear trajectory solution to the TPBVP and accounting
for statistical variations in the boundary conditions. The
performance function random variable under this treat-
ment is found to have a quadratic form, and Pearson’s
Approximation is used to model it as a chi-squared random
variable. Stochastic dominance is borrowed from mathe-
matical finance and is used to rank statistical distributions
in a metric sense. Analytical results and approximations
are validated and an example of the approach utility is
given. Finally conclusions and future work are discussed.
I. I NTRODUCTION AND BACKGROUND
In recent years there has been a significant increase
in trackable on-orbit objects due to new launches, col-
lisions, and improved sensor capabilities [1], [2], [3].
Object correlation for objects with maneuver execution
capability has become increasingly complex, and oper-
ational methods to correlate objects have become more
important, specifically for collision avoidance operations
[4], [5].
Object track correlation has an extensive body of
literature, particularly in continuous visible and radar
tracking applications. Many approaches use statistical
properties of sensors or known target qualities to mini-
mize false alarms and the effects of clutter. Probabilistic
Multi-Hypothesis Tracking (PMHT), Probabilistic Data
Association Filter (PDAF), and Modified Gain Extended
Kalman Filter (MGEKF) are representative of these
algorithms (discussions and examples of their usage are
in [6], [7], [8], [9]). Largely, these approaches update
their correlations as new measurements are generated.
An alternate class of problems to examine are those
with large gaps in observation, such as on-orbit object
tracking. In these scenarios the problem is to associate
individual object tracks incorporating 5-15 minutes of
observations (perhaps generated using some of the meth-
ods mentioned above) separated by observation gaps on
the order of tens of minutes to days [10], [11].
One way to support candidate object pairing associ-
ation is to propagate the initial track uncertainty (also
introducing process noise) and compute the Kullback-
Leibler Distance (KL-D) [12], the Battacharya Distance
(B-D) [13], or the Mahalanobis Distance (M-D) [14]
from the newly generated object track state and uncer-
tainty.
Alternately, with an initial and final track for each
candidate association pairing, each association may be
addressed as a Two Point Boundary Value Problem
(TPBVP) linking uncertain boundary conditions. In this
approach the optimal connecting trajectory performance
distribution may itself be used as a metric, and mutu-
ally exclusive track association pairings may be ranked
against one another in a metric sense. Specifically, LQR-
type costs (quadratic in state or fuel deviations from
a nominal homogeneous trajectory) directly measure
the effect of active maneuvers. Note that for space
applications, a subclass of this problem concerned only
with quadratic cost in fuel usage can be used [15].
This paper examines the full LQR performance index
to expand upon previous efforts. This proposed method
is fundamentally different from computing the statistical
KL-D, B-D, or M-D from the objects’ expected state
distribution as it directly accounts for control usage.
Computing the distribution of the LQR performance
index has the additional advantage that the resulting
probability distribution function may be used to support
hypothesis testing to infer intentions or detect maneu-
vers.
The contribution of this paper is to generalize from
quadratic control costs to full LQR trajectory costs.
Effort is made to maintain applicability to general
dynamical systems with LQR trajectory costs. Theory
2011 American Control Conference
on O'Farrell Street, San Francisco, CA, USA
June 29 - July 01, 2011
978-1-4577-0079-8/11/$26.00 ©2011 AACC 913