MIT OpenCourseWare

http://ocw.mit.edu

16.323 Principles of Optimal Control

Spring 2008

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.
 16.323 Lecture 13

LQG Robustness

• Stengel Chapter 6
• Question: how well do the large gain and phase margins discussed for LQR (6–29)
map over to LQG?
Spr 2008 16.323 13–1
LQG

• When we use the combination of an optimal estimator and an optimal

regulator to design the controller, the compensator is called
Linear Quadratic Gaussian (LQG)
– Special case of the controllers that can be designed using the sep­
aration principle.

• The great news about an LQG design is that stability of the closed-loop
system is guaranteed.
– The designer is freed from having to perform any detailed mechanics
- the entire process is fast and can be automated.

• So the designer can focus on the “performance” related issues, being

confident that the LQG design will produce a controller that stabilizes
the system.
– How to specify the state cost function (i.e. selecting z = Cz x) and
what values of Rzz, Ruu to use.
– Determine how the process and sensor noise enter into the system
and what their relative sizes are (i.e. select Rww & Rvv )

• This sounds great – so what is the catch??

• The remaining issue is that sometimes the controllers designed using

these state-space tools are very sensitive to errors in the knowledge of
the model.
– i.e., the compensator might work very well if the plant gain α = 1,
but be unstable if it is α = 0.9 or α = 1.1.
– LQG is also prone to plant–pole/compensator–zero cancelation,
which tends to be sensitive to modeling errors.
– J. Doyle, ”Guaranteed Margins for LQG Regulators”, IEEE Trans­
actions on Automatic Control, Vol. 23, No. 4, pp. 756-757, 1978.

June 18, 2008

T,(s)if

Spr 2008
if and 16.323 13–2

Excerpt from document by John Doyle. Removed due to copyright restrictions.

June 18, 2008

Spr 2008 16.323 13–3

• The good news is that the state-space techniques will give you a con­
troller very easily.
– You should use the time saved to verify that the one you
designed is a “good” controller.

• There are, of course, different definitions of what makes a controller

good, but one important criterion is whether there is a reasonable
chance that it would work on the real system as well as it
does in Matlab. ⇒ Robustness.
– The controller must be able to tolerate some modeling error, be­
cause our models in Matlab are typically inaccurate.
� Linearized model
� Some parameters poorly known
� Ignores some higher frequency dynamics

• Need to develop tools that will give us some insight on how well a
controller can tolerate modeling errors.

June 18, 2008

Spr 2008 LQG Example 16.323 13–4

• Consider the “cart on a stick” system, with the dynamics as given in

the following pages. Define
� � � �
θ q
q= , x=
x q̇
Then with y = x
ẋ = Ax + Buu
y = Cy x

• For the parameters given in the notes, the system has an unstable pole
at +5.6 and one at s = 0. There are plant zeros at ±5.

• Very simple LQG design - main result is fairly independent of the choice
of the weighting matrices.

• The resulting compensator is unstable (+23!!)

– This is somewhat expected. (why?)
4
10
Plant G
Compensator Gc
2
10
Mag

0
10

−2
10

−4
10
−2 −1 0 1 2
10 10 10 10 10
Freq (rad/sec)

200
Plant G
Compensator Gc
150
Phase (deg)

100

50

0
−2 −1 0 1 2
10 10 10 10 10
Freq (rad/sec)

Figure 13.1: Plant and Controller

June 18, 2008

Spr 2008 16.323 13-5

Example: cart with an inverted pendulum.

m, l • Nonlinear equations of motion can be developed for

θ large angle motion (see 30-32)
L
F

M
• Force actuator, θ sensor
x

Figure by MIT OpenCourseWare.

Linearize for small θ

(I+mL2) - mgL θ = mL

(M+m) +g -mL =F

⎡ ( I + mL2 ) s 2 − mgL −mLs 2 ⎤ ⎡Θ( s ) ⎤ ⎡0 ⎤

⎢ ⎥⎢ ⎥=⎢ ⎥
⎣ − mLs 2
( M + m) s + Gs ⎦ ⎣ x( s ) ⎦ ⎣ F ( s ) ⎦
2

Θ mLs 2
=
F ⎡⎣( I + mL2 ) s 2 − mgL ⎤⎦ ⎡⎣(M + m)s 2 + Gs ⎦⎤ − (mLs 2 ) 2

Cannot say too much more

Let M= 0.5, m=0.2, G=0.1, I=0.006, L=0.3

Θ 4.54s 2
Î gives =
F s 4 + 0.1818s 3 − 31.18s 2 − 4.45s

therefore has an unstable pole (as expected)

s= ± 5.6,-0.14,0

June 18, 2008

 Spr 2008 16.323 13-6

10

8 Near Origin

2
Imag Axis

-2

-4

-6

-8

-10
-6 -4 -2 0 2 4 6 8
Real Axis

Figure by MIT OpenCourseWare.

June 18, 2008

 Spr 2008 16.323 13–7

2
10
Loop L
1
10
Mag

0
10

−1
10

−2
10
−2 −1 0 1 2
10 10 10 10 10
Freq (rad/sec)

−50

−100
Phase (deg)

−150

−200

−250

−300
−2 −1 0 1 2
10 10 10 10 10
Freq (rad/sec)
Figure 13.2: Loop and Margins

June 18, 2008

Spr 2008 16.323 13–8

Root Locus

10
10
0.64 0.5 0.34 0.16

0.76 8
8

6
6 0.86

4
4
0.94

2
2
Imaginary Axis

0.985

0.985
−2
2

0.94
−4
4

0.86
−6
6

−8
0.76 8

0.64 0.5 0.34 0.16

−10
−10 −8 −6 −4 −2 100 2 4 6 8 10
Real Axis

Figure 13.3: Root Locus with frozen compensator dynamics. Shows sensi­

tivity to overall gain – symbols are a gain of [0.995:.0001:1.005].

Root Locus

2
2 0.64 0.5 0.34 0.16
1.75
0.76
1.5
1.5
1.25
0.86
1
1
0.75
0.94
0.5
0.5
Imaginary Axis

0.985 0.25

0.985 0.25
−0.5
0.5
0.94
0.75
−1
1
0.86
1.25
−1.5
1.5
0.76
1.75
−2 0.64 0.5 0.34 0.16
2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Real Axis

June 18, 2008

Spr 2008 16.323 13–9

• Looking at both the Loop TF plots and the root locus, it is clear this
system is stable with a gain of 1, but
– Unstable for a gain of 1 ± � and/or a slight change in the system
phase (possibly due to some unmodeled delays)
– Very limited chance that this would work on the real system.

• Of course, this is an extreme example and not all systems are like this,
but you must analyze to determine what robustness margins your
controller really has.

• Question: what analysis tools should we use?

June 18, 2008

Spr 2008 16.323 13–10
Analysis Tools to Use?
• Eigenvalues give a definite answer on the stability (or not) of the
closed-loop system.
– Problem is that it is very hard to predict where the closed-loop poles
will go as a function of errors in the plant model.

• Consider the case were the model of the system is

ẋ = A0x + Bu
– Controller
� also based on A0�, so nominal
� closed-loop dynamics:
�
A0 −BK A0 − BK BK
⇒
LC A0 − BK − LC 0 A0 − LC

• But what if the actual system has dynamics

ẋ = (A0 + ΔA)x + Bu
– Then perturbed closed-loop�system
� � dynamics are: �
A0 + ΔA −BK A0 + ΔA − BK BK
⇒
LC A0 − BK − LC ΔA A0 − LC

• Transformed A¯cl not upper-block triangular, so perturbed closed-loop

eigenvalues are NOT the union of regulator & estimator poles.
– Can find the closed-loop poles for a specific ΔA, but
– Hard to predict change in location of closed-loop poles for a range
of possible modeling errors.

June 18, 2008

Spr 2008
Frequency Domain Tests 16.323 13–11
• Frequency domain stability tests provide further insights on the sta­
bility margins.

• Recall from the Nyquist Stability Theorem:

– If the loop transfer function L(s) has P poles in the RHP s-plane
(and lims→∞ L(s) is a constant), then for closed-loop stability, the
locus of L(jω) for ω ∈ (−∞, ∞) must encircle the critical point
(−1, 0) P times in the counterclockwise direction [Ogata 528].
– This provides a binary measure of stability, or not.

• Can use “closeness” of L(s) to the critical point as a measure of

“closeness” to changing the number of encirclements.
– Premise is that the system is stable for the nominal system
⇒ has the right number of encirclements.

• Goal of the robustness test is to see if the possible perturbations to

our system model (due to modeling errors) can change the number
of encirclements
– In this case, say that the perturbations can destabilize the system.

June 18, 2008

Spr 2008 16.323 13–12

stable OL
1.5

0.5 LA(jω)
Imag Part

ω
0
2

−0.5 ω1 LN(jω)

−1

−1.5

−1.5 −1 −0.5 0 0.5 1 1.5

Real Part
Figure 13.4: Plot of Loop TF LN (jω) = GN (jω)Gc (jω) and perturbation (ω1 → ω2 )
that changes the number of encirclements.

• Model error in frequency range ω1 ≤ ω ≤ ω2 causes a change in the

number of encirclements of the critical point (−1, 0)
– Nominal closed-loop system stable LN (s) = GN (s)Gc(s)
– Actual closed-loop system unstable LA(s) = GA(s)Gc(s)

• Bottom line: Large model errors when LN ≈ −1 are very dangerous.

June 18, 2008

Spr 2008 16.323 13–13

Nichols: Unstable Open−loop System

1
10
Mag

0
10

−1
10
−260 −240 −220 −200 −180 −160 −140 −120 −100
Phase (deg)

Nichols: Unstable Open−loop System

1.05
1
0.99
1.04 1.01

1.03

1.02

1.01
Mag

0.99

0.98

0.97

0.96

0.95
−180.5 −180 −179.5 −179 −178.5
Phase (deg)

Figure 13.5: Nichols Plot (|L((jω))| vs. arg L((jω))) for the cart example
which clearly shows the sensitivity to the overall gain and/or phase lag.
June 18, 2008
Spr 2008 16.323 13–14
Frequency Domain Test
stable OL

0.5
Imag Part

|d(jω)|
|LN(jω)|
−0.5

−1

−1.5 −1 −0.5 0 0.5 1

Real Part
Figure 13.6: Geometric interpretation from Nyquist Plot of Loop TF.

• |d(jω)| measures distance of nominal Nyquist locus to critical point.

• But vector addition gives −1 + d(jω) = LN (jω)

⇒ d(jω) = 1 + LN (jω)

• Actually more convenient to plot

1 1
= � |S(jω)|
|d(jω)| |1 + LN (jω)|
the magnitude of the sensitivity transfer function S(s).
June 18, 2008
Spr 2008 16.323 13–15

• So high sensitivity corresponds to LN (jω) being very close to the

critical point.
Sensitivity Plot
3
10
|S|
|L|

2
10

1
10
Mag

0
10

−1
10

−2
10
−2 −1 0 1 2
10 10 10 10 10
Freq (rad/sec)

Figure 13.7: Sensitivity plot of the cart problem.

• Ideally you would want the sensitivity to be much lower than this.
– Same as saying that you want L(jω) to be far from the critical
point.
– Difficulty in this example is that the open-loop system is unstable,
so L(jω) must encircle the critical point ⇒ hard for L(jω) to get
too far away from the critical point.

June 18, 2008

Spr 2008 16.323 13–16
Summary
• LQG gives you a great way to design a controller for the nominal
system.

• But there are no guarantees about the stability/performance if the

actual system is slightly different.
– Basic analysis tool is the Sensitivity Plot

• No obvious ways to tailor the specification of the LQG controller to

improve any lack of robustness
– Apart from the obvious “lower the controller bandwidth” approach.
– And sometimes you need the bandwidth just to stabilize the system.

• Very hard to include additional robustness constraints into LQG

– See my Ph.D. thesis in 1992.

• Other tools have been developed that allow you to directly shape the
sensitivity plot |S(jω)|
– Called H∞ and µ

• Good news: Lack of robustness is something you should look for,

but it is not always an issue.

June 18, 2008