XM 2012 » Towards Early Emergent Property Understanding

Georg Hackenberg, Lehrstuhl für Software & Systems Engineering, Technische Universität München

Towards Early Emergent Property UnderstandingMerging Behavior Space Exploration and Model-Based Software Engineering

Extreme Modeling Workshop at MODELS 2012


The context of our contribution.

Model-based techniques are essential.

– From early validation to critical verification

Systems are becoming more complex.

– From systems to systems-of-systems

Effects arise from primitive interactions.

– From local decisions to global impact

Requirements constrain emergent properties.

– From behavior limitations to interaction limitations

Towards Early Emergent Property Understanding 2


An example for illustration.

Information Technology: Managing usage

Electric Power Grid: Constraining usage



The focus of our contribution.


Requirements arespecified on global behavior

Global behavioremerges from localbehavior

Local behavior cannotbe specified in earlyphases

▷ Model global behaviorwith respect to localbehavior

▷ Model requirementsboth on local andglobal behavior

▷ Model dominancerelationship betweenbehaviors

Problem Solution


Some related work in the field.


Software & systems modeling

UML [OMG]

– Rich but semi-formal

FOCUS [Broy]

– Formal but limited

Model checking

Bounded [NuSMV]

– Error behavior

Probabilistic [PRISM]

– Statistical indicators

Behavior space exploration

Robot planning

– Dynamic programming

Distributed control

– Model-predictive control


The big picture.


Behavior Space Exploration

Valid Goal-Oriented System Behavior

Non-Deterministic System Model Annotations

FOCUS System Theory 𝛿1

2

3

4


The core system modeling theory.


𝑜2.2

𝑜2.1

𝑜1.2

𝑜1.1 𝑜1.1(0) 𝑜1.1(1) 𝑜1.1(2) …

𝑜1.2(0) 𝑜1.2(1) …

𝑜2.1(0) …

…

𝑐1

𝑐2


The extensions to the modeling theory.


(Non-)Determinism

𝑜: 𝑇𝑖𝑚𝑒 → 𝐷𝑜𝑚𝑎𝑖𝑛

𝑜 𝑡 = 𝑒𝐷𝑜𝑚𝑎𝑖𝑛(𝑡)

Annotations

require

– Boolean observations

equal

– All observations

minimize/maximize

– Ordered observations

cost

– Minimized observations

Tmp

Cmd Egy

Prc

Bnd

Cst

minimize

requireequal


The behavior space exploration algorithm.

For transition from time step 𝑡 to 𝑡 + ∆𝑡:

1. Generate: Choose non-deterministic options

2. Calculate: Derive deterministic variables

3. Verify: Require boolean observations

4. Prune: Determine dominant behavior

5. Sort: Prioritize remaining behavior

non-deterministic

deterministic

require

equal/mini-/maximize

cost



The comparisons for pruning behavior.

Obervation Annotation Comparison

𝑜1 equal 𝑜1𝐴 𝑡 = 𝑜1

𝐵 𝑡

…

𝑜𝑛+1 minimize 𝑜𝑛+1𝐴 𝑡 ≤ 𝑜𝑛+1

𝐵 𝑡

…

𝑜𝑛+𝑚+1 maximize 𝑜𝑛+𝑚+1𝐴 𝑡 ≥ 𝑜𝑛+𝑚+1

𝐵 𝑡

…



A visualization of the search space (1/3).

startoff

on

invalid

valid









The case study for demonstration.


Energy domain

Volatile producers

– Sun or wind

Smart prosumers

– Fridge or storage

Energy autonomy

– Use local energy

System variant one

System variant two

+

+ +

Local Power Grid

Global Power Grid

𝑐1 … 𝑐𝑘

minimize flow


The model for system variant one.


Sun

– Power = Gaussian

Refrigerator

– Command = On/off

– Power = 0/-200 Watt

– Temperature = Rise/fall

– Constraint = Min/max

Model

– Balance = Power difference

– Cost = Balance integral


The model for system variant two.



The performance of exploration.



The selected decision strategies.


System variant two

Phase 1: Delay cooling

Phase 2: Load storage

Phase 3: Unload storage

System variant one


Phase 2: Cool down



A reflection on the annotation scheme.

Autonomy as cost minimization

– Low overall balance deviation

Frige temperature as equivalence class

– High temperature better if much energy expected

– Low temperature better if few energy expected

Frige temperature interval as boolean constraint

– Hard lower and upper limits

Storage level minimium as boolean constraint

– Hard lower limit, no upper limit

Storage level as value maximization



The proposed approach in retrospect.

Model defines non-deterministic behavior

– What are possible local decisions?

– What are local decision effects?

Annotations define behavioral dominance

– What behavior do I consider to be valid?

– What behavior do I consider to be better?

Algorithm explores behavioral alternatives



The positive and the negative.


Model fundamentals

– Phyical effects

– Cost formulation

Annotate dominance

– Behavior constraints

– Equality classes

– Multi-dimensional order

Explore behavior

– Generic algorithm

Limited expressions

– Primitive types

– Basic operations

Limited models

– Discrete time

– Static structure

Limited performance

– Space explosion

(-) Negative(+) Positive


The conclusions of the study.

Current state:

The annotations improve goal understanding

The idea works well for „easy“ problems

The selected strategies improve confidence

The algorithm complexity increases quickly

The notions of system/dominance are limited

Future work:

Working with larger case studies

Introducing probabilistic behavior terms

Experimenting with learning-based approaches



You can visit us at …


http://smartgrid.in.tum.de/

http://smartgrid.in.tum.de/


The end.

Thank you for your attention!
