IEIIT-CNR SWAN, ARRI ©RT 2006 1 Las Vegas and Monte Carlo Randomized Algorithms for Systems and...
-
Upload
stephen-watkins -
Category
Documents
-
view
214 -
download
0
Transcript of IEIIT-CNR SWAN, ARRI ©RT 2006 1 Las Vegas and Monte Carlo Randomized Algorithms for Systems and...
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
1
Las Vegas and Monte Carlo Randomized Algorithms for Systems and Control
Roberto Tempo
IEIIT-CNR
Politecnico di Torino
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
2
Overview
A Success Story Randomized Algorithms, Monte Carlo and Las Vegas Some Recent Research Directions Applications: High Speed Networks and UAV
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
3
A Success Story
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
4
A Success Story
Randomized AlgorithmsRandomized Algorithms (RAs) are successfully used in
various areas, including computer science, numerical
analysis, optimization, …
… but in systems and control their use is often limited
to Monte Carlo simulations
Example: Example: Sorting problem
Algorithm:Algorithm: RandQuickSort (RQS)
RQS is implemented in the Linux sorting command
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
5
RandQuickSort (RQS)
given n real x1 x2 x3 need to sort them
numbers x4 x5 x6 in increasing order RQS is an iterative algorithm consisting of two phases 1. randomly select a number xi (e.g. x4)
2. perform deterministic comparisons between xi and (n-1) remaining numbers
x2 x3 x4 x1 x5
x6
numbers smaller than x4 numbers larger than x4
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
6
Running Time of RQS
Because of randomization, running time may be different from one execution of the algorithm to the next one
RQS is very fast: average running time is O(n log (n)) This is a major improvement compared to brute force
approach for example when n = 2m
Average running time is also a highly probable running time (Chernoff bound)
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
7
Randomized Algorithms, Monte Carlo and Las Vegas
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
8
Randomized Algorithm: Definition
Randomized AlgorithmRandomized Algorithm (RA): An algorithm that makes random choices during execution to produce a result
For hybrid systems, “random choices” could be switching between different states or logical operations
For uncertain systems, “random choices” require (vector or matrix) random sample generation
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
9
Monte Carlo and Las Vegas RA
Monte Carlo Randomized AlgorithmMonte Carlo Randomized Algorithm (MCRA): A randomized algorithm that may produce incorrect results, but with bounded error probability
Las Vegas Randomized AlgorithmLas Vegas Randomized Algorithm (LVRA): A randomized algorithm that always produces correct results, the only variation from one run to another is the running time
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
10
Uncertain Systems
Consider random uncertainty and a bounding set B is a (real or complex) random vector (parametric
uncertainty) or matrix (nonparametric uncertainty) Consider a performance function
J( Rn,m → R
and level > 0 Define worst case and average performance
Jmax max J(JaveEJ(B
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
11
Example - H Performance
H performance of sensitivity function
S(s,) = 1/(1 + P(s,) C(s))
J() = ||S(s,)||
Objective:Objective: Check if
Jmax andJave
These are uncertain decision problems
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
12
Two Problem Instances
We have two problem instances for worst case performance
Jmax and Jmax >
and two problem instances for average case performance
Jave and Jave >
This leads to one-sidedone-sided and two-sidedtwo-sided MC randomized algorithms
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
13
One-Sided MCRA
One-sided MCRA:One-sided MCRA: Always provide a correct solution in one of the instances (they may provide a wrong solution in the other instance)
Consider the empirical maximum
JJmax max J(i
where N is the sample size
Check if JJmax or JJmax >
i1,…,N
^
^ ^
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
14
One-Sided MCRA: Case 1
Jmax
JJmax
JJmaxJmax <
algorithm provides a correct solution
1 2 3 4 56
J(1)
J(2)
J(3)
J(4)
J(5)
J(6)
J
^
^
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
15
One-Sided MCRA: Case 2
JmaxJJmaxmax
Jmax >
JJmax<
algorithm may provide a wrong solution
1 2 3 4 56
J(1)
J(2)
J(5)
J(4)
J(6)
J(3)
J
^
^
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
16
Two-Sided MCRA
Two-sided MCRA:Two-sided MCRA: They may provide a wrong solution in both instances
Consider the empirical average
JJave ave J(i
where N is the sample size
Check if JJave orJJave >
i1,…,N
^
^ ^
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
17
Two-Sided MCRA
Jave
1 2 3 4 56
J(1)
J(2)
J(3)
J(4)
J(5)
J(6)
J
JJave^
Jave >
JJave< ^
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
18
Two-Sided MCRA
Jave
1 2 3 4 56
J(1)
J(2)
J(3)
J(4)
J(5)
J(6)
J
JJave^
^Jave <
JJave>
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
19
Las Vegas Randomized Algorithms
We also have zero-sidedzero-sided (Las Vegas) randomized algorithms
Las Vegas Randomized Algorithm (LVRA):Las Vegas Randomized Algorithm (LVRA): Always give the correct solution
Running time may be different from one run to another LVRA has more limited applicability than MCRA
Example:Example: RandQuickSort
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
20
Current Research on LVRA
Switched systems:Switched systems:
- design a common Lyapunov function for systems
x(t) = A x(t)
where A is an interval matrix with entries ranging between upper/lower bounds
Consensus control:Consensus control:
- design randomized algorithms achieving finite-time average consensus for connected networks
.
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
21
Uncertain Systems, Optimization, System Identification
From common to piecewise Lyapunov functions[1]
Ellipsoidal randomized algorithm[2] and stopping rules[3]
RAs for semi-infinite programming[4]
MRAS methods for global optimization[5]
Estimation via MCMC[6]
RAs for model validation[7] and system identification[8]
…
[1] H. Ishii, T. Basar and R. Tempo (2005) [2] S. Kanev, B. De Schutter and M. Verhaegen (2002) [3] Y. Oishi and H. Kimura (2003) [4] V. B. Tadic, S. P. Meyn and R. Tempo (2006) [5] J. Hu, M.C. Fu and S.I. Marcus (2005) [6] J.C. Spall (2004)
[7] M. Sznaier, C. M. Lagoa and M.C. Mazzaro (2005) [8] X. Bombois, G. Scorletti, M. Gevers, P. Van den Hof and R. Hildebrand (2006)
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
22
Applications of RAs
RAs have been developed for many control applications
Control of flexible structures
Robustness of high speed networks
Stability of quantized sampled-data systems
Control design for brushless DC motors
Synthesis of real time embedded controllers
Mini-UAV control design
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
23
Applications of RAs
RAs have been developed for many control applications
Control of flexible structures
Robustness of high speed networks
Stability of quantized sampled-data systems
Control design for brushless DC motors
Synthesis of real time embedded controllers
Mini-UAV control design
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
24
Mini-UAV Control Design
Study and development of a
real-time land control and
monitoring system for fire
prevention in Sicily
Uncertainty description
Development of three RAs for
gain synthesis and robustness
analysis (according to flying
quality military specs)
[1] L. Lorefice, B. Pralio and R. Tempo (2006)
IEIIT-CNR
SWAN, ARRI SWAN, ARRI ©RT 2006©RT 2006
25
References
“Randomized Algorithms for Analysis and Control of
Uncertain Systems” by R. Tempo, G. Calafiore and F.
Dabbene, Springer-Verlag, 2005
Additional documents, papers, MATLABTM
codes, etc, please consult
http://staff.polito.it/roberto.tempo