Upload
others
View
6
Download
0
Embed Size (px)
Citation preview
Fluidization of�Discrete Event Modelsor a�marriage between the discrete
and�the continuous
Manuel SilvaUniversidad�de�Zaragoza
1. As�an�appetizer…�on�models,�systems�and�fluidization…
2.�(Discrete)�Petri�Nets:�elements�to�be�used
3.�Fluidization�of�autonomous�discrete�models
4.�Fluid�timed�models:�server�semantics…
Informal, trying to provide intuitionUsing examples as simple as possible
Well known,�model is:*�in�Science &�Technology: the representation*�in�Fine�Arts:�the reference…
…�just the contrary.
La�trahison des�images,René�Magritte,�1928Ͳ29
…�but if model is a representation,�it is only “an image”, “a view”
constructed with a�purpose,�not “the reality”:a
A MODEL vs THE (REAL)�SYSTEM
how representative is a�model?�ͲͲͲ Fidelity vs complexity?
This system is discrete or continuous?
1)Continuous?2)Discrete?3)…???
The system &�the model:
1) Purpose2) Fidelity3) Complexity
Models are “views”
…�sometimes in�manufacturing:• a�production engineer?• or and�hydraulic engineer?
• Flows• Levels (deposits…)• 3Ͳways�valves(mixing flows)
• …
In�manufacturing,��������in�informatics…������most frequently:rendez-vous lead�to�a�minimum operator
In�population dynamics,�in�chemistry…�are�mostfrequent other functions
The state explosion problem…�(�a�very naive view…)�H
omot
ecy
Increasing the population (or initial marking):����������������������• the approximation improves• much bigger computational savings
MarkovianT-time
interpretation
SteadyState
FLUIDThe bigger the marking,�the better theaproximation &�the computational savings
If we fluidify (from integer to�reals)��the model&�use�infinite server�semantics (min�operator)…infinite
From individuals to�populations:decoulourization
A�classical &�basic predatorͲprey system (revisited)
If bounded populations
Coloured PN Place/Transition net (PN)
VolterraͲLotka
If we fluidify (from integer to�reals)��the previous model&�use��product server�semantics…
The timed continuousPN�is Live &�Bounded
The timed discrete PN�isNonͲlive &�Unbounded
LotkaͲVolterra orbit
but…
product
If populations are�as�in�the LotkaͲVolterra model:
?
Fluidization of DEDS (also “views”…)*�Queuing�Networks�(Newell,�1971)*�Petri�Nets�(David&Alla:�Continuous�PNs,�1987�//��
Silva&Colom,�PNs�&�LPP,�1987+1999)*�Process�Algebra�(Hillston,�2005)
Fluid “views” of DEDS*�Forrester (or�Stock�&�Flow)�diagrams (1961)*�Stochastic Flow (or Fluid) Systems (Cassandras,���
2001) “We�use�Stochastic�Fluid�Models�(SFM)�for�control�andoptimization�of�communication�networks�in�whichdetailed�discrete�event�models�become�impractical.”
Fluidization and�fluid�“views”�(to�overcome the state explosion problem) With a�preliminary character…
If populations (i.e.,�initial markings)�are:• “veryͲvery big”:�fluid�views are�most surely very interesting• “big”:���fluid�views may usually be�very interesting• “not so�big”:��fluid�views may or may not be�
(very)�interesting
* k >>> 1* k >> 1* k >~ 1
With fluidization (total�or partial):• What we keep?• What we lost?• What we gain?
The balance?¾ Like in�differential equations:�NOT�all can�be�
approximated by linearization (e.g.�Lorentz)¾ In�PNs:�NOT�all can�be�approximated by
(partial)�fluidization
How to��improve the (partially)�fluid�relaxations?
1. As�an�appetizer2.�(Discrete)�Petri�Nets:�elements�to�be�used¾ A�quick�global�perspective�of�place/transition�nets�(P/T)¾ Fundamental�equation�and�mathematical�programming¾ Components,�invariants�&�stable�predicates¾ Removing�spurious�solutions�&�implicit�places¾ Rank�theorems:�a�proof�of�synergy�between�the�
performance�and�logic�analysis
3.�Fluidization�of�autonomous�discrete�models
4.�Fluid�timed�models:�server�semantics…
Place/Transition nets�(P/T,�PNs)
PNs as�directedbipartitemultigraph
PNs and�incidencematrices
Dynamicsystem
Marking:*�Distributed*�Numerical
Marking (numerical distributed state)�& evolution
Thenegation
is notpresent!
Logic:consumption/ production……….…………..
No mono-tonicity
OR constructions AND constructions
Marking (numerical distributed state)�& evolution
Synchronizations (two mechanisms…):*�Joins /�RendezͲVous*�Weigths on arcs
Fundamental�or state (transition)�equationWithm�=�m0 +�C�V,�two main computationalproblems:1)�in�the naturals (computational cost)�&2)�with spurious (validity of�computations…)
Our focus is today on:1)�Fluidization:�relaxation into the nonͲnegative reals2) How to�remove spurious solutions on:
2.1)�The discretemodel:�untimed &�timed2.2)�The fluidmodel:�untimed &�timed
…�in�the PN�paradigm
Fundamental�equation &�fluid�relaxation
Decomposed views…�(I):�Invariants
From the fundamental equation (structure):�1.�Annuler vectors of�the incidence matrix,�C:* Left &�nonͲnegative:�PͲsemiflows (y):��y C = 0, y �0*�Rigth &�nonͲnegative:�TͲsemiflows (x):�C x = 0, x �02.�Invariants (laws…):* y m0 =�y m�ͲͲͲ token conservation law (PͲinv.)* Firing &�marking repetitive sequences law (TͲinv.)3.�Subnets:�*�Conservative component (PͲsubnet)*�Repetitive component (TͲsubnet) Duality
The semiflows allows to�“see” relevantcomponents of�themodel (some “intrinsic”�descompositions)
y C = 0y �0
� A single�(elem.)�TͲinvariant:�X�=�(1�1�1�1�1�1�1�1�1�1)� In�this case,�the TͲsubnet is the full�net�(i.e.,�no�decomposition)
Decomposed views…�(I):�Invariants
The 4�elementary components (3�FS&DA�+�1�store/SM)*�Proof of�properties*�Synthesis of�controllers*�Implementation (programmed,�PͲprogrammed,�hardwired…)
Decompose to�analyze,�synthetize &�implement
Generators of elementary traps (also siphons) can be computed as P-semiflows of a transformed net (join work MASI/LIP6-UNIZAR, 93)
Decomposed views…�(II):�stable predicatesLooking the net�as�a�directed &�bipartite graph (structure):1.�Inclusions concerning subsets of�places (here we do�notconsider weigths on arcs…):* Trap: subset of�places�s.t.�the subset of�output�transitionsare�included in�their input�subset of�transitions*�Siphon: the inverse of�a�trap2.�Stable predicates (laws…):* Trap:�if marked,�cannot be�fully unmarked (tokens are�trapped)* Siphon:�if emptied,�cannot be�marked again3.�Subnets:�*�Trap component (PͲsubnet)*�Siphon component (PͲsubnet)
Reverseconcepts
Decomposed views…�(II):�stable predicates� p1 is a trap &�a�siphon,�but not a�conservative
component
Conservative components are�trap &�siphon components� y =�(1�0�1)�is a�PͲsemiflow,�thus:�m1+m3 =10� y =�(0�1�1)�is a�PͲsemiflow,�thus:�m2+m3 =11
How�to�cut�the�spurious�deadlock�m�=�(0�1�10)?
� m3+q3 =�9,�
q3:�complement�of�m3
� p1 is�an�initially�marked�trap,�
thus:�m1 ш�1
� but:�m1+m3 =�10,�thus
m2+m3 =�11�,�thus�m1 ш�2��(p2 has�2�frozen�tokens)
(m1�
Removing aspurioussolution thatunmark a�trapis a�polynoͲmial time�problem
2�frozentokens
m3+q3 =�9
m3 ч�9
q3 makes�p2 implicit
Implicit placesPlaces�comptabilize theresources:�constraint thefiring of�transitions
Sufficient condition in�polynomial time�(PPL)
Implicit place:�never is theunique to�constraint…�but:¾ Interleaving semantic?:�
Seq.�IP¾ Step semantics?:�Con.�IP
• Elimination:�to�simplify the implementations;�to�reduce�in�analysis(boundedness,�liveness,�reversibility,�etc.;…
• Addition:�to�create new�traps (&�PͲsemiflows);�to�remove spurioussolutions;�to�allow a�nonͲintrinsic decomposition;�to�increasethe Hamming distance;…
Implicit places,�spurious solutions &�traps
� The addition of�implicit places�p,�r &�s remove all the natural�spurious solutions.
� In�general:�Alleviate,�but do�not dress (always)
Removing natural�solutions is beyond the integer hulldefined by the fundamental�equation (…�Gomory’s cuts)
m1+m2+
+m4+ms=�1
� Did siphons also help to�cut spurious deadlocks?�:�YES!
A GMEC
• A�live S3PR�model• S=�{r2,�r3,�q1,�q4,�p1,�p4}�
is a�siphon• m,�is a�solution of�the
fundamental�equation:¾ m[r1]�=�m[q3]=�
m[p3]�=�1,¾ &�zero for the rest
• m empties S.• Thusm is a�killing
spurious solution• The addition of�place�u
remove the killingspurious solution
Avoiding siphons to�be�emptied,�killing (spurious)�solutions can�be�removed�
Remark:�The added place�u is implicit in�the discretePN�system (makes the system DF�as�continuous)
Implicit places,�spurious solutions &�siphons
A GMEC:m[u]+m[q3]+m[p3]=1
Logical &�performance�properties…
Improvement1: Adding the implicit place p6, a new P-component is createdImprovement2: embeded QNs with product form (GNQN…)
* Descomposition in�PͲcomponents* Solvable�by�a�LPP (polynomial�time)* Idea:�detection�of�bottlenecks 2:�Throughput bound of�t5?
1:�Are�p2 &�p3 in�mutex?
(adding p6)
From performance�evaluation:�a�family of�rank theorems
… found when computing the visit ratios:
• Necessary condition for structural boundedness &�liveness• N&S�cond. for SB&�SL�in�net�subclasses (e.g.,�equal conflict,�DSSP…)• Sufficient condition for structural boundedness &�liveness
Polyn.�time+�Synergy
Str.�non�live ¿?
SB & SL: ¾Consistent:
C x = 0, x � 1¾Conservative:
y C = 0, y � 1¾ rank(C)ч�/ECS/Ͳ1
rank (C)�=�4rank(C)ч�4Ͳ1�=�3?
� Coherence� Economy� Synergy
The Petri�nets�paradigm
+�Fluid�PNs+�Hybrid PNs+�…
1. As�an�appetizer2.�(Discrete)�Petri�Nets:�elements�to�be�used3.�Fluidization�of�autonomous�discrete�models
3.1�Fluidization of�DES�&�fluidizability3.2.�Basic�properties�of�autonomous�continuous�
models3.3.�Improving�the�approximation�(I):�
removing�spurious�solutions
4.�Fluid�timed�models:�server�semantics…
3.1�Fluidization of�DES:�concept�&�fluidizability
Definition:��A�continuous�transition*�Enabling degree:
*�Firing:����
Neighbour places�become�continuous
¿¾½
¯®
0>][][
][min)enab(t, p,tp,tp Pre
Premm
m mD t� o� '
Continuous net:��all transitions are�continuousHybrid net:��onlysome transitionsare�continuous
Reachability (under total�continuity)
*�m�is reachable fromm0 iff a�finite sequence
exists,�such that*�Reachability space,�����������������=��set�of��reachable
markings
–
kktt DD V �11
mm0 ��� o���� o�DD kktt
11 �
), RS(N 0m
),(NRS ), (NRS CD 00 mm �
00.2
0.40.6
0.81
00.2
0.40.6
0.81
0
0.5
1
m(p1)m(p2)
m(p
3)
Same reachability space
All (discrete)�PN�models can�be�fluidified?
ZENO
nonͲlive as�discrete &live as�continuous
live as�discrete &�nonͲlive as�continuous
¾ At�the limit:�a�trap is emptied¾ So�traps are�no�more�traps…!�
DeadlockͲfreeness:�discrete vs continuous (!!!)
Live:�yes�or not?�
New�concept: limͲreachability
A�main question:�When a�given property is preserved by fluidization?Fluidization:�can�be�approximated by the scaling of�the initial marking¾ homotheticmarkings!¾ amonotonicity property
Monotonicities
DeadlockͲfreefor the marking,�and
Bigger markingand�deadlock!
Homothetic monotonous,�…but not monotonous
x k
3.2. Basic�properties�of�autonomous�continuous�models
Let�us�assume�in�the�sequel�that:1)�The�net�N�is�consistent�(��x�>�0,�Cͼx =�0 )�and2)�all�the�transitions�can�be�fired�“a�little”�at�m0
(otherwise�stated:�no�siphon�is�empty�at�m0),
RS(N,�m0)�=�{m|m =�m0�+�CͼV,��V t 0}* Thus RS(N,�m0) is linearly delimited,�it is a�convex set* Spurious�solutions?:�
• NOT�on�the�continuousmodel;• But�spurious�solutions�in�the�originally�discrete net�model�are�here�integrated.�¾ This�raise�the�question�of�how�to�remove�it!�
Marking homotecy and�DeadlockͲfreeness
(y C = 0, y >0)
Sufficient condition for DF…
Marking homotecy and�properties preservation
23 = 8 LS
Approaching DF�(N is consistent & all siphons initially marked)
Needed?
k =�1
Approaching DF�(N is consistent & all siphons initially marked)� Structural (upper)�bound of�p:�SB(p)�=�{maxm(p) /�m�=�m0 +�C�V`� Structural lower bound (frozen tokens)�of�p:�SF(p)�=�{maxm(p)�/�
/�m�=�m0 +�C�V`
1 LS+�4�var.�+�2�equ.
¾ Firing languages are�identical if the samem0 and�tp &�tq are�“silent”�transitions¾ No�system has�a�a solution,�thus they are�deadlock-free systems
k =�1
8 LS
1 LS
3.3.�Improving�the�autonomous (thus,�in�general,�all�timed)�approximations�:�cutting�spurious�solutions
Complementary (efficient)�techniques:
� Enforce the bound to�the bigger integer of�¾ the marking of�places�(addition of�complementary places)¾ the enabling of�transitions (addition of�selfͲloop places)
� …� … below Gomory cutting planes
Truncating
� Spurious solutions that empty a�trap (like in�discretemodels;�reachable in�the limit)� Spurious solutions that empty a�siphon (feasible withfinite sequences)
if integer … beyond Gomory cutting planes
If the initial marking is “not too big”,�any SB(p)�&�SF(p)�should be�truncated to�natural�values:�floor functionsSB(p3)�=�1,5:�m(p3)+m(q3)�=�1¾ Adding q3 avoids siphonS�=�{p1,�p4}�to�be�emptied in�the continuous approximation
Adding q4 does not prevent the non�integer (deadlock)�solution:�
m =�(0�0�0�1,5�1,5)Thus the integer hull is not obtained(too expensive for the benefits…)
Decreasing�improvement�with�increasing�markings...
Remarks:�*�Non�homotheticvalue:�NOT�9k!*�Relative changedecrease with k
m�=�(0�0�10)�is�a��spuriousdeadlock�
Removing a�knownspurious solutionbecause a�trap isemptied is a
polynomial problem
1. As�an�appetizer2.�(Discrete)�Petri�Nets:�elements�to�be�used3.�Fluidization�of�autonomous�discrete�models4.�Fluid�timed�models:�semantics,�properties
and�improvements4.1�(Deterministic)�Infinite�server�(&�product)�semantics4.2�Patterns�of�behaviour4.3�Improving�the�timed�fluidization�(II)
4.3.1�Bound�reaching�problem�and�roͲsemantics4.3.2�Stochastic�approximation
x
m�=�m0 +�C�V (autonomous:�state transition eq.)
… deriving
m�W��=�C�f�W�� and��m(0)�=�m0,�with f�W��=V�W��x
… if the firing is a function of time
¾ How is f defined?�o firing semantics? Exist several…¾ An interesting source of�inspiration isMarkovian PNs
(inheritance of�results,�coherence with discrete perforͲmance models):�lead�to�infinite server semantics
4.�Timed�continuous�models:�semantics,properties�and�improvements
m�W��=�m0 +�C�V�W�
Infinite server�and�product (population) semantics
In�the sequel wejust concentrate
on infiniteserver�semantics
ʄ�=�[0.5�0.1�1.0�0.3]
From configuration to�operation mode
From the graph (structure):�1.�Configuration (allocation):�A�set�of�pͲt arcs defining a�1Ͳcover of�all transitions:�those that constraint the firing…2.�Region:�the subͲreachability space defined by a�configuration (are�disjoint except on the boundary)3.�Operation mode:�the dynamic linear�system corresponͲding to�a configuration
4.1.�(Deterministic)�Infinite server�semantics
3�configurations out of�4�possible for ʄ�=�[0.5 0.1 1.0�0.3] &m0 =�(1�1�0�0)
1)�Configuration,2)�Partition (except
on the borders)�of�thepolytope &��������3)�Phase portrait
Region 4�is notreached in�the
trajectory
m3 ч�m4 &�m2 чm3
m3 шm4 &m2 чm3
m3 ч�m4 &m2 ш�m3
m3 шm4 &m2 ш�m3
About “infinite servers�semantics”1. The�firing�flow�is�proportional�to�the�input�“level”2. The�differential�system�is:
1. Ordinary2. Piecewise�(the�minimum operator)�linear�with�constant�
coefficients�(idea�of�configuration)3. BUT:�pͲflows�(the�marking�of�some�places�are�NOT�state�
variables,�but�constraints)�lead�to�PIECEWISE�AFFINE3. Flows are bilinear on firing speeds and�markings:�it reflects
the duality of transitions & places.4. Continuous�PNs�are�technically�hybrid�(the�discrete�state,�
the�configuration,�is�implicit�in�the�continuous)5. Under�infinite�server�semantics:�Time�differentiable6. BUT:�able�to�simulate�TURING�MACHINES!!!�(ENS�Cachan /�
UNIZAR,�2007)
Infinite servers�semantics.�Properties
� Positive�systems:�f(W)�t 0,�m(W)�t 0���Marking &�time�homothecies (duality):� If m’(0)�=�k ͼm(0)�thenm’(W)�=�k ͼm(W)
&�f�’(W)�=�k ͼf(W)� If O’ =�k ͼO thenm’(W)�=�m(k ͼW)
&�f�’(W)�=�k ͼf(k ͼW)
But no�superposition,�what lead�to�many counterͲintuitive behaviors
Counterintuitive�properties (I)
• The�steadyͲstate of�the�throughput�of�the�continuous�PN�relaxation�may�not be�an�upper�bound�of�corresponding�discrete�models
2
p3
2
5
4
p1
p4
p5p
2
t1 t2
t3 t4
Discrete: 0.801 f/t.u.
Continuous: 0.535 f/t.u.
To fluidify does not improve!
Counterintuitive�properties (II)
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
p3
2
5p1
p4
p5p
2
t1 t2
t3 t4
• No�performance�monotonicities with respect toresources
m[p5]=3 � F[t3]=1.071
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Counterintuitive�properties (II)
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
p3
2
5p1
p4
p5p
2
t1 t2
t3 t4 m[p5]=4 � F[t3]=0.535
• No�performance�monotonicities with respect toresources
m[p5]=3 � F[t3]=1.071
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
.Discretemodel
Counterintuitive�properties (II)
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
p3
2
5p1
p4
p5p
2
t1 t2
t3 t4 m[p5]=4 � F[t3]=0.535
m[p5]=5 � Deadlock
• No�performance�monotonicities with respect toresources
m[p5]=3 � F[t3]=1.071
Counterintuitive�properties (III):�steadyͲstate� No performance monotonicities with respect to firing speeds� Discontinuity / Bifurcation (loss of hyperbolicity)
Discrete model
Performance�monotonicity� MonoͲTͲsemiflow reducible (MͲTͲSͲred)�nets�are�conserͲvative with a�single�TͲsemiflow if Homothetic EqualConflicts [Pre�(t1)�=�q�Pre�(t2)�]�are�reduced.
� Let N be�a�MͲTͲSͲred�net.�Assume that steadyͲstates are�reached for given firing speeds and�initial markings.�If thestͲst configurations contains the support of�a�PͲsemiflow,� <�N,�O,�m1>�&�<N,�O,�m2>,�with m1�ч�m2�,��or� <�N,�O1,�m0>�&�<N,�O2,�m0>,�with O1�ч�O2
then f1 ч�f2��(i.e.,�system 2�never slower!)
Duality
� Corollary:�SL&SB�EQ�net�systems are�throughputmonotonous wrt themarking and�the firing speeds. m(p1)�+�m(p3)�=�10 m(p1)�+�m(p3)�=�10,
only if m(p1)�=�5!No�PͲcomponent“covered by theconfiguration”
Monotonousoperation mode
Non�monotonousoperation mode
M-T-S net 3 of 4 conf.
4.2.�Patterns�of�behaviours
(1) Equilibrium (point attractor)
p4
p3
p2
p1
p6
p5
p7
t1
t2
t3
t4
t5t6
32
4
2
2
2)�Basic�oscillatory (linear)
H : hiringF : firing
E : employeesP : productionS : sales
St : stock
Patterns�of�behaviours
3)�Oscillatory (nonͲlinear)�&�timing depending collapse:�Siphons Flows and�phaseͲportraits:�collapse if O2 >�10
O2 >�10
O2 <�1
Two discontinuities wrt O2 (�at�1.0�&�10.0) High�&�low frequencies and�collapse
+++ Small�variations in�parameters:¾ Orbits (like in�LotkaͲVolterra predatorͲprey model),¾ Limit cycles¾ …�
2
k
t1 t2
t3 D
~r ~f
r f
20
2020D��
4.3�Improving�the�timed fluidization�(II)With a�preliminary character:��use�all improveͲ
ments for untimed models that remove spurioussolutions!
In�particular,�those that :1)�empty a�marked trap2)�empty a�siphon (and�are�known to�be�spurious)3)�add complementary places�(qi)�that truncate the structural boundof�places
Also:�4)�add selfͲloop places�that truncate the enabling boundof�transitions
Reuse…:Interleavingqualitative & quantitative
sf = spurious-free(i.e., m[p2]=9
(k = 1)
Removing the spuriousdeadlockm=(0�1�10)T
O� O� O� SPN TCPN TCPN+q3 TCPN+f11 3,00 1,00 1,00 0,480 0,0062 0,500 0,00632 0,40 1,00 10,00 0,346 0,476 0,385 0,400
If O1 (vs O2)�is relatively very fast:1)�The basic TCPN�approach thespurious deadlock on the discretemodel.2)�The addition of�q3 (complementof�p3)�remove the spuriousdeadlock,�leading to�theremarkable improvement in�theTCPN+q3 approximation.
f1 may help to�improve,�but hereit is improbable…
Adding marking and�flow truncating placesBecause close to�a�spurious deadlock
4.3.1 Bound reaching problem & U-semantics
� What to�do�whenm0 is “not that big”,�and�exist transiͲtions with enabling degree =�1�(particularly if a�synchroͲnization appears as�weight on the arc from p�to�t)???
� One possibility:�to�mantain the transition t1 as�discrete(problem similar�to�responseͲtime in�RC�circuits).
� Alternative IDEA:�To�divide�the load�of�tokens in�such a�way that the firing of�the transition start with a�treshold value:�k–U¾ Question:�a�“good”�value for U ?�(heuristic?)
Infinite server�semantics Hybrid net UͲsemantics
The enabling degree of�t1
Infinite server�semantics
Hybrid approximation
With anheuristic
technique to�choose thevalue for U
4.3.2 Stochastic approximation
�If the discrete PN�model to�be�approximated is providedwith exponential timing (Markovian PN,�analyzable bymeans of�a�Markov Chain that is isomorphous to�thereachability graph),�then:
The fluid�PN�approximation lead�to�a�system ofstochastic piecewise affine differencial equations:�like the deterministic +�white gaussian noise
� m0 “very big”:�applicable a�kind of�functional law of�numbers or deterministic fluid�limit
� m0 “not that big”: applicable a�kind of�functionalcentral�limit theorem (in�the line�of�Donsker theorem)
Stochastic Continuous Petri�Nets�(STCPN):�adding Gaussian White�Noise to�the firing flow
Noise:independentnormallydistributed RVswith average
and�covariancematrix
STCPN: obtainedadding thenoise to�thefiring flows
STCPNs:�particularly interesting when “moving”�among two or more�configurations
Exp {max}�ш�max {Exp}�
Improvement when the discrete system in�the steadyͲstate switch among different regions
Transient for m0
m�=�m0 +�C�V is:1)�in�the naturals (computational cost)�&2)�with spurious (care with computations …)
Our focus today,�was on:1)�Fluidization:�relaxation into the nonͲnegative reals2)�How to�remove spurious solutions on:
The discrete and�corresponding fluidmodels(untimed &�timed)
Integratedview in�the PN�paradigm
Concluding remarks
GOAL:��To�overcome the state explosion problem…�in�certain cases�(the constraint:�the existence of�nonͲfluidizable PN�models)¾ The bigger the marking: the better the approximation &��������
the bigger the saving on computations.¾ Useful with untimed &�timed net�models¾ Reuse /�adaptation of�results from discrete PNs theory:�
Integration in�the PN�paradigm (economy;�coherence;�synergy)¾ Partial fluid�relaxations:�(a�kind of)�Hybrid Petri�Nets
Concluding remarks
The use�of�“good”�timed fluid�(technically hybrid)�models:� Parametric optimization (�where to�go?,�dim.�of�buffers?...�� Sensibility analysis
� Observers
� Controllers (min.�time;�…)� Diagnosis�under faulty behaviours� …
“Good”�is related to:�� the relative absence of�spurious solutions� the level of�fluidization:�Hybrid vs continuous nets?� In�timed models:�an appropriate server�semantics
(deterministic or stochastic)?
StructuralGenericPunctual
Fluidization of�Discrete Event Modelsor a�marriage between the discrete
and�the continuous
Manuel SilvaUniversidad�de�Zaragoza