Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus,...
Transcript of Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus,...
![Page 1: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/1.jpg)
Morphisms: Exact and Approximate
Bernard ZeiglerProf. Emeritus, Univ. Arizona
Chief Scientist, RTSync
Journees DEVS Francophone
Cargese, Corsica
April 13, 2016
![Page 2: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/2.jpg)
Outline• The need to consider multi-level model approximation families
• Some background theory – dynamic systems and morphisms
provide a mathematical basis
• Some sufficient conditions that enable exact approximation
• Approximate Morphisms to handle departures from enabling
conditions
• Error propagation and example of sensitivity analysis
• Construction of multi-level model approximation families
![Page 3: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/3.jpg)
The need to consider multi-level model approximation families
• Systems of Systems (SoS) may be modeled at different levels of spatial extent and resolution (municipal, county, state…)
• Models may be constructed at coarse abstraction levels that support faster simulation - e.g., model checking – but not Pacman
• Models may be constructed at high fidelity that are purported to be closer to reality, but how well do we know the details
• How can we ensure that errors introduced in the aggregation/disaggregation processes lie within acceptable limits?
• Interoperation of models at different levels of resolution presupposes effective ways to develop and correlate the underlying abstractions
• An useful methodology for integrated family of approximation models would allow
• flexibility in calibration and cross-calibration
• application to diverse experimental frames, i.e. properties
![Page 4: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/4.jpg)
Implementation: Case Study:A Watershed behavior
Specialization for two abstraction WS models
Specialization for two abstraction Basin models
2Montain models
![Page 5: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/5.jpg)
Implementation: Results
On what basis should they agree?
=> Comparing simulated
and observed daily flows
for the one-year period
(365 days for the year
2009).
What if they disagree?
Results of simulation of 4 pruned and transformed compositions: Each figure presents a time series
![Page 6: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/6.jpg)
Some background theory – dynamic systems and morphisms provide a mathematical basis
• Systems theory generalizes finite state and linear theory exact and approximate model approximation
• Framework for M&S provides needed concepts for approximate model construction, including model-simulator separation and experimental frames
• Systems theory provides morphism concepts for exact model simplification
• M&S Framework enables System morphism concepts to be extended to handle approximate model construction and error propagation analysis
![Page 7: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/7.jpg)
Stratification of System Knowledge/Specification Levels
![Page 8: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/8.jpg)
Associated Structure/Behavior Preserving Morphisms
![Page 9: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/9.jpg)
Modeling and Simulation Framework
Source System Simulator
Model
Experimental Frame
SimulationRelation
ModelingRelation
behavior database
BASE Model
LUMPED Model
Exact and Approximate Morphisms
Abstraction is defined as valid model simplification and is relative to one or more experimental frames
PLOS Computational Biology: Minimum Information About a Simulation Experiment (MIASE)
![Page 10: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/10.jpg)
Experimental Frames Determine Validity of Abstraction
= BASE
= LUMPED
mapping
state
stateSame output
i.e. mapping
Rrefines the
Output partition
![Page 11: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/11.jpg)
Multi-step morphism allows micro states and input/output encoding/decoding
![Page 12: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/12.jpg)
Some sufficient conditions that enable exact abstraction
• Finite state theory provides algorithms employing congruence relations, state set partitions, and output refinement
• Linear systems theory developed model realization theory that was shown to be essentially equivalent to finite state theory
• Probabilistic automata provided somewhat different lumpability criteria
• These concepts were generalized in dynamic system theory within the M&S framework
![Page 13: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/13.jpg)
State system homomorphism: exact and approximate
Homomorphism concepts borrowed from finite state automata
require base and lumped model states to remain in state correspondence over time
Exact homomorphism: zero error
= BASE
= LUMPED
mapping
state
state
statestate
transitionstate
mapping
state
Transition introduced error
transition
transition
mapping
error after twotransitions
state
state
state
Approximate: have to consider error accumulation over time
![Page 14: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/14.jpg)
state mapping
parameter
mapping
parameters
parameters
state
state
Parameter Morphism: a homomorphism that maps the
parameter space of the base model into that of the lumped
model
BASE Model
LUMPED Model
Parameter Morphisms retain valid simplification und er changes in parameter valuesThis supports cross-calibration in multi-level, mu lti-resolution families
![Page 15: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/15.jpg)
System Theory Generalization of Exact Aggregation Conditions
a)
b) c)
componentaggregation
couplingaggregation
baseblock
lumpedblock
ab
c a cd
Theorem: Sufficient conditions for Aggregation
Let there be partition on the components of a base model for which 1) blocks have
homogeneous components, 2) the coupling indifference conditions hold for all pairs of blocks
and 3) the base model output is computable from block model census counts.
Then a lumped model exists for which there is an exact homomorphism based on an census
count aggregation.
The Aggregation theorem states that
not only must there be homogeneity
within blocks but also the inter-block
element-to-element interactions must
satisfy coupling indifference conditions
![Page 16: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/16.jpg)
Example: Lumpability criteria applied to linear discrete time systems
• The base model transition function is expressed as a matrix of coefficients
• The lumped model transition function is of lower order with coefficients derived from aggregation of base coefficients
• The state mapping aggregates base states to lumped states
• Lumpability criteria provide sufficient conditions for preservation of the state mapping over time
1 11 21 31 41 1
2 12 22 32 42 2*
3 13 23 33 43 3
4 14 24 34 44 4
x a a a a x
x a a a a x
x a a a a x
x a a a a x
′ ′ =
′ ′
1 11 21 1*
2 12 22 2
y b b y
y b b y
′ = ′
1 0.5 * ( 1 2)
2 0.5 * ( 3 4)
y x x
y x x
= += +
0,5( 31 41 42 44)a a a a+ + +
Lumped model
Base model
Parameter
morphism
State mapping
31 41
32 42
a a
a a
==
Lumpability
criteria
![Page 17: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/17.jpg)
• Approximate system morphisms loosen up the strict requirements of exact
system morphisms
• Propagation of error that results is characterized and related to the
dynamics of the resulting lumped model
• The error propagation can grow or attenuate with time depending on
lumped model stability characteristics
Approximate Morphisms handle departures from
enabling conditions
• We can
– analyze a base model and a block partition construction for error departure
from exact morphism conditions
– simulate and compare base and lumped model behavior for sensitivity to
error propagation
![Page 18: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/18.jpg)
Example of error propagation analysis: Approximate Lumpability
0 0
0 0 0 0
0,5 0,5 0 0 0
0,5 0,5 0 0
g e g
e g
g g g
g g
−
→
g e error at time 5
error behavior
0.8 0.001
-0.002 error decreases over time
0.8 0.01
-0.02 decreases
0.8 0.1 -0.2 decreases
0.8 0.0 0.0 constant
1 0.1 0.4 error increases over time
1.2 0.1 7.0 increases explosively
2 2
( )
( 1 2 1) ( 3 4 2)
error t
x x y x x y
=
+ − + + −Departure from
lumpability = e
Theory predicts that the gain, g determines
error propagation characteristics:
• Error decreases over time when g<1
Otherwise it increases
• The closer to lumpability, the faster the
error disappears
![Page 19: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/19.jpg)
Application to evolving model approximation families
• Approximate model construction methodology supports integrated
families of approximate models
• Approximate models are indexed by the experimental frames in which
their conditions approximately hold – error propagation conditions in
these frames can be estimated
• As field data is ingested for V&V, calibration of any one model propagates
information to others from which it is derived and which are further
approximations of it, thus maintaining a coherent model family
integrated families of approximation models support 1) estimating trade-offs in accuracy and
execution time and 2) choice of approximations meeting analysis needs and time constraints.
![Page 20: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/20.jpg)
Preservation/Predictive Ability (“predictivity”) of models
• Preservation: Does the lumped model preserve a given property of the base model?
• Predictivity: Does a given property of the lumped model imply that the property holds for the base model?
• Example: Recurrent (cyclic) vs Absorbing (acyclic) behavior
Base
model
Lumped
model
morphism
![Page 21: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/21.jpg)
All base models
Systems
Not Having P
Systems
Having P
All lumped models
Models
Having P
Models
Not Having P
Lyndon’s theorem for automata:
S has positive P implies M has P
M has negative P implies S has P
Deceptive case:
Possible but may be
rare
:
• Conditions for inheritance of stability properties for continuoussystems (Foo)
• Downward preservation of p implies upward preservation of –p
• Properties of interest seem not to be expressible in negative form.,
• Sierocki applied Lyndon's theorem to finite automata.
• Reachability, connectedness, and reversibility and are positive bydirect statement.
• Sierocki shows that upward inheritance of positive properties holds forthe usual homomorphism of automata.
• Moreover, downward inheritance holds for negative properties.
• Saadawi and Wainer showed that some properties flow upward from
safety timed automata models verified in uppaal to real time advance
devs models under a strong form of bi-simulation similar to
isomorphism.
a positive property is one expressible in first order logic without use of negation.
a negative property is one that requires negation to express it.
![Page 22: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/22.jpg)
Preservation/Predictive Ability (“predictivity”) of Markov models from analysis of their underlying Directed Graphs(DG)
* Theorem If C is a directed cycle, then G hom→ C iff G contains only cycles of length divisible by the length of C
Pavol Hell, Huishan Zhou, Xuding Zhu Homomorphisms to oriented cycles. 2003
• Sequences (DAG) can map to DAGs and to Cycles (with low probability)
• Cycles (DCG) can map to only DCGs - a cycle either maps to a single
state (if it is all in an equivalence class) or to a proper cycle*
So
• Lumped Model cycles can only come from Base Model cycles
• Lumped Model sequences come from sequences with high probability
• So (Property Preservation)
• Base Model is recurrent implies Lumped model is recurrent
• Base Model is absorbing implies Lumped model is probably absorbing
• And (Property Predictivity)
• Lumped Model is recurrent implies Base model is probably recurrent
• Lumped Model is absorbing implies Base model is absorbing
Connected Acyclic DG
Digraph
homomorphisms
Strongly
Connected DG
![Page 23: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/23.jpg)
DEVS makes it easy to cross deterministic /stochastic lines
DEVS
Deterministic
DEVS
(FDDEVS)
Stochastic
DEVS
(FPDEVS, Markov)
Deterministic
model
Markov
model P = 2/3
Count the number of transition from
First block to Second
Probability assigned to
transition
Markov
model P = 2/3
Extends to Lumping of
Markov Models
morphism
![Page 24: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/24.jpg)
Generated 10,000 pairs of randomized stochastic matrices and state partitions
• The square matrices were sampled from a uniform distribution of sizes, m in the range [4, 34] with partitions sampled conditionally on m with uniform distribution in the range [2,m-1].
• Matrices were generated as sequences of m*m entries sampled with Binomial distribution with probability increased in 10 steps from .1 to 1 each with 1000 trials (with normalization to assure row summation to unity).
• Partitions were generated by partitioning the integer m, selecting one randomly, and randomly permuting the order of the partition numbers
• For each (matrix, partition) pair, the lumped matrix was generated using Equation 1 (Definition 2) so that the original and lumped chains are called the base, lumped model pair.
![Page 25: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/25.jpg)
Frequencies of Base/Lumped Cycle Numbers using Cycle Detection in Jama package
0 0 38
0 1 147
1 0 28
1 1 9569
1 2 2
2 0 4
2 1 180
2 2 10
3 1 19
3 2 1
4 1 2
LumpedBase Number
![Page 26: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/26.jpg)
Lumping usually reduces the number of transitions:Avg Outdegree of Lumped is usually less than Base
0
5
10
15
20
25
30
35
1
13
3
26
5
39
7
52
9
66
1
79
3
92
5
10
57
11
89
13
21
14
53
15
85
17
17
18
49
19
81
21
13
22
45
23
77
25
09
26
41
27
73
Avg Outdeg Base Avg Outdeg Lump
![Page 27: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/27.jpg)
However, Predictive uncertainty is increased by Lumping: Trajectory Entropy of Lumped usually larger than that of Base
0.00
0.20
0.40
0.60
0.80
1.00
1.201
95
18
9
28
3
37
7
47
1
56
5
65
9
75
3
84
7
94
1
10
35
11
29
12
23
13
17
14
11
15
05
15
99
16
93
17
87
18
81
19
75
BaseTrajEntropy LumpedTrajEntropy
![Page 28: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/28.jpg)
Lumpability example taken from Wikipedia
1 2 3 4
1 2
Partition mapsStates 3 ,4 into
state 2Partition mapsStates 1,3 into
state 1
Morphism mapsalong row
into their sum
Note these sums
are equal
Sums along rows = 1
Prob of transition Of state 1 to state 2
This morphism is preserved under
Single step transitions
the lumped model has the same
long run behavior as the base
![Page 29: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/29.jpg)
But what about this matrix?It still is stochastic but does not satisfy lumpability
1 2 3 4
1 2
- eps+ eps
However it is close To the original so shouldn’t Its behavior also be close?
- eps/2+ eps/2
![Page 30: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/30.jpg)
Lumpability metric: LumpSTD measures how varied the partial sums are – the more varied the less lumpable
1 2 3 4
1 2
Consider the sums assamples of a rvCompute the sample mean and stdThe mapped value = sample meanLumpSTD be the std.
![Page 31: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/31.jpg)
Jensen-Shannon Metric For Probability Distributions
JSDist = square root (JSD) is a distance metric
Approximation Error =
JSDist (Lumped Equilibrium Distribution,
Mapped (Base Equilibrium Distribution))
![Page 32: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/32.jpg)
Background values of JSDist metric
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1 2 3 4 5 6 7 8 9
JSDist increase with Distribution Size
Dist between Rand & Rand Dist between Rand & Unif
Dist between Unif & Certain
Error should beLess than this levelTo claim it is better
than random
![Page 33: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/33.jpg)
Error propagation is self-damping: general theory says Contraction mapping implies error decreases
0.00
0.01
0.01
0.02
0.02
0.03
0.03
0.04
0.04
0 10 20 30 40 50 60 70 80 90 100
Contraction MaxRatio
= max Do/Di in one transition neighborhood
Di
Do
Contraction ratio is always <1
Single Steptransition
![Page 34: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/34.jpg)
Approximate Morphisms: What’s the probability of finding a reasonably good approximate lumpablepartition when sampling at random?
Base
Lumped
3 % of space isNear exact
Lumping
![Page 35: Morphisms: Exact and ApproximateMorphisms: Exact and Approximate Bernard Zeigler Prof. Emeritus, Univ. Arizona Chief Scientist, RTSync JourneesDEVS Francophone Cargese, Corsica April](https://reader033.fdocuments.in/reader033/viewer/2022051921/600ea47902d4827f9244f965/html5/thumbnails/35.jpg)
Conclusion
• Properties of interest may not be preserved in the abstraction process
• Moreover lumped model properties may be deceptive – they may not carry over to the base model
• Measures of model quality may not behave as expected under aggregation
• Exact morphisms are very rare
• But approximate morphisms may be much more common and discoverable by modelers – humans and computers