Symbolic dynamics of Markov chains

Report
Symbolic dynamics of Markov
chains
P S Thiagarajan
School of Computing
National University of Singapore
Joint work with:
Manindra Agrawal, S Akshay, Blaise Genest
Acknowledgements
• Samarjit and Javier
• IAS
• TÜV SÜD Foundation
Agenda
• Study complex dynamical systems
– verification of temporal logic specification
• High dimensional
• Continuous time; continuous value domains.
– Differential equations
– Hybrid automata
• Networks of such dynamical systems
Strategy
• Approximate the dynamics.
– Discretize the time and value domains
• Today’s first talk will have this flavor.
– Sample the dynamics and do a statistical analysis.
• Today’s second talk
• To handle large networks: Deterministic synchronizations
• Tomorrow’s talk
Probabilistic dynamics
• Illustrate these ideas in the setting of probabilistic
dynamical systems.
– Yield useful approximations in the presence intractability and
lack of knowledge.
– Scalability can be achieved through sampling (simulations)
based statistical analysis techniques.
• Statistical model checking
• Key application: (Networks) of biochemical networks.
Study I: Symbolic dynamics of Markov
chains
•
•
•
•
Well-known probabilistic dynamical systems.
Rich theory
Widely applied.
Our focus (finite state, discrete time):
– Symbolic dynamics
• Discretize the probability value space [0, 1].
Symbolic dynamics of Markov chains
• Analyze the symbolic dynamics via:
– classical (linear) temporal logic specifications.
• Probabilities are sneaked in through the atomic propositions.
– Model checking methods
• Details in:
– LICS’2012
– Journal paper under review.
Two views of Markov chains
• View I
– a finite state probabilistic transition system.
• View II
– a linear transform of probability distributions over the states.
The transition system view
1
3
1
2
1
1
2/5
3/5
4
The transition system view
0
1
1
3
1
2
3/5
1
1
0
2/5
4
0
The transition system view
3
1
3
1
1
1
2
1
4
3/5
2/5
1
2/5
4
2/5
3/5
1
2
3
3
4
4
2/5
3/5
The transition system view
• Outcomes: Infinite paths
• Basic cylinder: The set of infinite paths that have a
common finite prefix.
• Path space: -algebra generated by the basic cylinders.
• Probabilities are assigned to basic cylinders in a natural
way.
• Extends canonically to a probability measure over the
path space.
The transition system view
3
1
3
1
1
1
2
4
3/5
2/5
1
2/5
1
4
1
Pr(B) = 1  2/5  1  1
= 2/5
2
1
1
B – The set of all paths that
have the prefix 3 4 1 3 4
3/5
3
3
1
1
4
B
4
Measurable properties
• The state 3 will be visited infinitely often.
PCTL, CTL, ….
Baier, C. and Katoen, J.-P. Principles of
Model Checking
VIEW II
• The Markov chain transforms (linearly) a given
distribution over its nodes to a new one.
• The graph of the chain is represented as a stochastic
matrix
– (all row sums are 1)
• The Perron-Frobinius theory (for non-negative matrices)
applies.
VIEW II
• Leads to natural sub-classes:
– Irreducible and aperiodic,
– periodic, …
• transient states, recurrent states….
• Stationary distributions….
VIEW II
1
3
1
2
1
1
2/5
3/5
4
The second view
0
1
3
1
2
1
1
1
0
2/5
3/5
4
0
Trajectory of distributions
• Markov chain as (linear)transformer of probability
distributions
3
1
1
1
2 0
3/5
1
1
0
2/5
4
0
Transition matrix M
Trajectory of distributions
0
0
3
1
1 1
1
2
3/5
1
2/5
0
4
Transition matrix M
• Each initial distribution generates a trajectory of distributions:
Trajectory of distributions
0
0
3
1
1 1
1
2
3/5
1
2/5
0
4
Transition matrix M
• Each initial distribution generates a trajectory of distributions:
Trajectory of distributions
• Dynamics:
– The set of trajectories generated from an initial set of
distributions
– This set can be infinite
• We can ask whether all (none) of the trajectories satisfy
some desired dynamical property
• What we can ask in this setting is incomparable with
View I specifiable properties
Trajectory of distributions
• Eventually the probability of being in 1 is always greater
than being in 3.
• There is a future time at which 90% of the probability
mass lies on {1, 4}.
Symbolic dynamics
• A trajectory:
– an infinite sequence of probability distributions (over a finite set
of nodes).
– The alphabet of probability distributions -over a finite set of
nodes- is (uncountably) infinite.
• Hence a trajectory is an infinite sequence over an infinite
alphabet,
• But exactly tracking probability distributions may be
neither necessary nor possible.
– Bio-chemical networks
– Sensor networks…
Symbolic dynamics
• We propose to reason about sequences of distributions
in a symbolic way.
– Using a finite alphabet of discretized distributions.
Symbolic dynamics
• A classical tradition dynamical systems
theory.
– Jacques Hadamard(1898), Morse and
Hedlund (1921), ….
– Shannon, Smale…
– Significant applications in data storage
and transmission (via coding theory),
linear algebra.
Symbolic dynamics
• We know how to do this for:
– Timed automata
– In some restricted settings for hybrid automata.
• For finite state Markov chains this has been done under
very restricted setting and unnatural restrictions (Chada
et.al 2011, Korthikanti et.all 2010.)
• We shall instead give up on bisimulation-based
equivalence classes:
• Instead, fix a discretization of [0, 1]
– With no restrictions handle all finite state Markov chains .
Symbolic dynamics
c
d
F2(x)
F(x)
F3(x)
e
a
b
x
Finite alphabet!
Symbolic dynamics
• Each block is a letter.
• Each trajectory is now recorded as a
sequence of letters taken from a finite
alphabet.
• If each block bisimualtion equivalence
class it is called a Markov partition!
• Study the system dynamics in terms of
these sequences.
– Sophic shift sequences
– Shift sequences of finite type.
Symbolic dynamics
• We know how to get “Markov partitions”) do this for:
– Timed automata
– For hybrid automata, in a few settings.
• For finite state Markov chains this has been done under
very restricted setting and unnatural restrictions (Chada
et.al 2011, Korthikanti et.all 2010.)
• We do not look for bisimulations of finite index.:
• Instead, we fix a discretization of [0, 1]
– Handle all finite state Markov chains ; no restrictions.
Symbolic dynamics of Markov chains
• Discretization
– We partition [0,1] into a finite set of intervals,
a
b
c
Symbolic dynamics of Markov chains
a
b
c
• Each (probability) value is mapped to
(identified with) the interval it falls in.
Symbolic dynamics of Markov chains

Symbolic dynamics of Markov chains
•
Each distribution  maps to D, a unique n-tuple of intervals.
•  ∈D means Γ() = D

= D
Symbolic dynamics of Markov chains
• Each distribution  maps to D, a unique n-tuple of intervals.
• Several distributions can map to the same D,
• in fact Γ-1(D) can be infinite

= D

Symbolic dynamics of Markov chains
• Discretization
– We partition [0,1] into a finite set of intervals,
• Γ-1(D) can be empty
?

= D
Symbolic dynamics of Markov chains
• Discretized distribution
– A tuple of intervals D is discretized distribution ( - distribution
for short ) if Γ-1(D) ≠ ø
ø

= D

= D ×
Symbolic dynamics of Markov chains
• The set of discretized distributions  is finite
Symbolic dynamics of Markov chains
• A Trajectory of M starting from a distribution  :
• Induces a symbolic trajectory
– a word over ω
• ξ  = (Γ() Γ(1) Γ(2) ….)
Symbolic dynamics of Markov chains
3
0.45
1
2
0.12
1
1
1
0.35
2/5
3/5
4
0.08
.

.
Symbolic dynamics of Markov chains
• Fix a set of initial distributions IN.
• We use a -distribution Din to specify IN. That is IN = Din
This can be an infinite set .
– IN =
=
,
………..
• We can fix the set of initial distributions in many other
ways.
Symbolic dynamics of Markov chains
• The symbolic dynamics of (M , Din ) is:
• We wish to reason about this -language.
Symbolic dynamics of Markov chains
• The discretization need not be uniform
• Can be different for each dimension (node).
• {[0, 1]} can be used to mask out “don’t care” nodes.
– Dimension-reduction.
A temporal logic to reason about the
symbolic dynamics
• I, a discretization of [0, 1] :
– < i , d > is an atomic proposition
– node i interval d
– “In the current distribution , (i) falls in the interval d ”.
– In the current discretized distribution D, D(i) = d
• Probabilistic linear-time temporal logic LTL I
The verification problem
• Given,
–
–
–
–
a Markov chain M,
a discretization I,
an initial set of distributions represented as Din ,
and a specification φ as an LTL I formula,
• Determine if M, Din ⊨ φ. In other words,
– (0) ⊨  for every  in LM
– Does every symbolic trajectory of M satisfy φ, i.e.,
– is it the case LM ⊆ L φ?
Example formulas
• Whenever probability of node i is “high”, the probability of
node j is “low” :
• The  -distribution (d1, d2, . . . , dn) appears infinitely often:
Example formulas
• Extending with FO theory of reals, we can express much more:
– e.g., Infinitely often, the probability of node i is at least twice the sum
of probabilities of all other nodes.
• Logics based on path spaces (PCTL etc.) and logics based
sequences of probability distributions are incomparable
Examples
• We have modeled a simple version of the Google
pageranking algorithms.
• We have also modeled a small pharmacokinetics system
for drug delivery.
The verification problem
• Given, M, I, Din , φ, determine if M, Din ⊨ φ , i.e., LM ⊆ L φ
– If LM is ω-regular and effectively computable then we can use standard
model checking techniques.
– But LM is not always ω-regular !
The verification problem
LM is not always ω-regular.
-approximations
• Fix a small  > 0 (fraction of length of an interval of I )
.

.
’
-approximations
• Fix a small  >0 (fraction of length of an interval of I )
.

D

.

.
’
-approximations
• Fix a small  >0(fraction of length of an interval of I )
.

D

.

.
’
• The -nbhd of a concrete distribution  is the set of all distributions which are atmost -away from it.
-approximations
• The -nbhd of a concrete distribution  is the set of all distributions which are atmost -away from it.
– Nbr()
•  is an -neighborhood iff there exists  such that
Nbr() = 
-approximations
Suppose D, D’ belong to a final class. Then there exist 
in D and ’ in D’ such that (, ’)  2
• Suppose D, D’ belong to a final class. Then for every 
in D and ’ in D’ , (, ’)  2 +  where  depends
only on the discretization.
•

D

DD
2
’ D’
DD
A basic property of the symbolic dynamics
• There exists a computable constant K and a computable
ordered collection of -nbhds { 0, 1, . . . , } called
final classes such that:
– After K steps, ξ  will forever cycle through the ordered final
classes. In other words, ξ (k) ∈ k mod  for k > K
A basic property of the symbolic dynamics
• There exists a computable constant K and an ordered
collection of -nbhds { 0, 1, . . . , } called final
classes such that:
– After K steps, ξ  will forever cycle through the ordered final
classes. In other words, ξ (k) ∈ k mod  for k > K

M
1
M
2
K steps
0
.

1
M
-1
M
M
2
A basic property of the symbolic dynamics
• There exists a computable constant K and an ordered
collection of -nbhds { 0, 1, . . . , } called final
classes such that:
– After K steps, ξ  will forever cycle through the ordered final
classes. In other words, ξ (k) ∈ k mod  for k > K

M
1
M
2
K steps
0
.
M
1
M
-1
M
M
2
-approximations of trajectories
• A symbolic trajectory ξ’ is an - approximation of ξ  (the
symbolic trajectory gen. by ) iff:
– ξ (k)= ξ’ (k) for all 0≤ k ≤ K ,
– For all k> K , ξ’ (k) and ξ (k) belong to the same final class.
Two approximate verification problems
• Determine if
– M, Din ⊨ φ , i.e.,
for every  ∈ Din , there exists an -approximation of ξ  in L φ.
– M, Din ⊨ φ , i.e.,
for every  ∈ Din , every -approximation of ξ  is in L φ.
Two approximate verification problems
• Proposition
– M, Din ⊨ φ implies LM ⊆ L φ
– M, Din ⊭ φ implies LM ⊈ L φ
Two approximate verification problems
• If M, Din ⊨ φ and M, Din ⊭ φ then M - approximately
meets φ.
• If we want to do better we can set ’ = /2 and iterate
(incremental overhead).
The main result
• Theorem:
Checking M, Din ⊨ φ and M, Din ⊨ φ are both
effectively solvable problems.
The main result
• The proof first assumes a single initial distributions and
proceeds to establish the theorem for:
– Irreducible aperiodic chains
– Irreducible periodic chains.
– General chains
• This is then extended to a set of initial distributions Din.
• The main step is the transient + steady state
characterization of the symbolic dynamics.
The main result
The main result
Additional results
• Convex hull of a finite set of distributions can also be
handled.
• The atomic propositions can be constraints expressed in
the first order theory of reals.
• Restrictions on the eigenvalues can yield -regular
behaviors (egs.: distinct and real)
Summary
• Discretizing the values space [0, 1] leads to an interesting
and useful symbolic dynamics of finite state Markov
chains.
• Extensions:
–
–
–
–
Optimizations
Implementations
Case studies
Interval Markov chains

similar documents