Physical models are important but difficult to build for adaptive cyber-physical systems (CPS). A single modeling formalism is not enough, and formalisms should be selected based on their expressiveness, supported analyses, and the relevant engineering expertise. Physical models should be obtained in a way that maximizes their value, analytical power, and fragility. Both theory-driven and data-driven approaches have tradeoffs. Finally, physical models should be treated as first-class entities in adaptation and be continuously adapted based on their predicted value and computational cost at runtime.
9. Abstractions of physical objects and
interactions
Beyond simple discrete models
Objects may be in the system, in the environment,
or on the border
Example: power model forTurtleBot
How much does each task consume?
How much power is left given current voltage?
How long does it take to charge?
9
10. Software models guide state-of-the-art
adaptive systems
Physical models are often implicit or assumed
In CPS, we need both software and physical
models!
10
11. 1. Selecting modeling formalism
2. Obtaining physical models
3. Using physical models in adaptation
11
12. 1. Selecting modeling formalism
2. Obtaining physical models
3. Using physical models in adaptation
12
13. Many formalisms and tools are available for
modeling CPS
Differential equations, signal flow graphs, automata
Position: no single formalism is enough to
model adaptive CPS; we need to embrace their
multiplicity
13
14. Evaluate individual formalisms
Expressiveness
▪ Linear/non-linear, continuous/discrete, classes of
functions (polynomials, transcendental functions, etc.)
Types of analyses supported
▪ Trade-off between expressiveness and computing cost
Engineering expertise
▪ Novices: higher effort and lower quality
We need approaches to integrate formalisms!
Difficult problem, outside of talk’s scope
14
15. We chose a linear real-valued regression model
Continuous changes in parameters
Easily embeddable into other models
15
P(v, t) = Av + Bt + C
15
time (s)
power(wh)
16. 1. Selecting modeling formalism
2. Obtaining physical models
3. Using physical models in adaptation
16
17. Goal: maximize value of each model
Analytical power: strength of predictions and
explanations
Fragility: amount of rework to accommodate future
changes
Computational cost: amount of processing needed
for analysis
Position: the way we build physical models
affects their value.We need more guidance!
17
18. Theory-driven
Physical theory
dictates first principles
Calibrate with data
18
Data-driven
Collect data first
Then create
abstractions from it
time (s)
power(wh)
19. 19
We chose to use data-driven approach
Low expertise with theory-driven models
Ok with low-precision far-horizon predictions
The model is fragile: hard to change
20. 1. Selecting modeling formalism
2. Obtaining physical models
3. Using physical models in adaptation
20
21. Software models in adaptation are used for:
State estimation and prediction
Triggering adaptive changes
Choosing adaptive strategy
+ Continuous improvement of models themselves
Position: physical models should also be
treated as first-class entities in adaptation
21
22. Clear representation
Either separate models or explicit embedding
Easier change and reuse
Coordinated use with cyber models
Estimation, prediction, choice
Models themselves should be adapted
Model value & cost should be the guiding factors
Need to reason about model value at run time!
22
23. Physical models in adaptive CPS are
important and difficult to build
23
Challenge Position
Selecting modeling
formalism
Embrace multiplicity; use formalisms
based on expressiveness, analyses, and
expertise.
Obtaining physical models Model value should the guiding factor.
More guidance is needed to connect
model- building and model value.
Using physical models in
adaptation
Physical models should be treated as first-
class entities and adapted based on their
value at run time.