from engineer to alchemist, there and back again: an alchemist … · 2013. 7. 9. · from engineer...
TRANSCRIPT
SAPERE Self-aware Pervasive Service Ecosystems
From Engineer to Alchemist, There and Back Again: AnAlchemist Tale
Danilo Pianini – [email protected]
Alma Mater Studiorum—Universita di Bologna
Cesena, May 24th, 2013
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 1 / 46
Outline Do or do not, there is no try.
1 SimulationDefinitionsModelsWhen it’s useful?
2 AlchemistSAPERE backgroundComputational modelEngine and architectureCase studiesPerformance
3 DevelopmentDevelopment modelContributorsFuture activities
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 2 / 46
Simulation
Simulation
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 3 / 46
Simulation Definitions
Scientific Method
Traditional science workflow [Parisi, 2001]
Traditional scientific method
identificationdirect observationtheories / hypothesisempirical observationquantitative analysisvalidation / invalidation
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 4 / 46
Simulation Definitions
Definition of Simulation
A new way for describing scientific theories
[Parisi, 2001]
Simulation is the process with which we can study the dynamicevolution of a model system, usually through computational tools
[Banks, 1999]
Simulation is the imitation of the operation of a real-world process orsystem over time
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 5 / 46
Simulation Models
Simulation Requires a Model
M. Minsky – Models, Minds, Machines
A model (M) for a system (S), and an experiment (E) is anything to whichE can be applied in order to answer questions about S.
Representation / abstraction
Formalisation
Aggregation, Simplification, Omission
Building a model...
How complex should be the model?
Which assumptions should be done?
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 6 / 46
Simulation Models
From Model to Simulation. . .
Computer simulation
Models are designed runnable processes
Simulation creates a virtual laboratory
controlled conditionsit’s easy to modify the experiment (variables, parameters, etc.)
Simulations imitate the operations of the modelled process
generation of an artificial evolution of the system
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 7 / 46
Simulation Models
. . . and Back
Deductions on the real system represented
Evaluation of theories about the model
Model validation [Klugl and Norling, 2006]
if the data is reliable;
if prediction doesn’t match the observed behaviour do not match
⇒ the model must be revised
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 8 / 46
Simulation When it’s useful?
Why do we Need Simulations?
[Parisi, 2001, Klugl and Norling, 2006]
The real system cannot actually be observed
The time scale of the real system is too small or too large forobservation
The original system doesn’t yet exist (or not anymore)
The system is too complex
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 9 / 46
Alchemist
Alchemist
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 10 / 46
Alchemist SAPERE background
The SAPERE Project
http://www.sapere-project.eu/
SAPERE (Self-aware Pervasive Service Ecosystems) is an EU STREPproject under the FP7 FET Proactive Initiative: Self-Awareness inAutonomic Systems (AWARENESS)
The objective of SAPERE is the development of a highly-innovativetheoretical and practical framework for the decentralized deploymentand execution of self-aware and adaptive services for future andemerging pervasive network scenarios
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 11 / 46
Alchemist SAPERE background
SAPERE World
LSA Space
LSA
LSALSA
LSA
eco-lawEngine
Software services
SAPERE Node
eco-law activation
SAPERENode
SAPERENode
SAPERENode
SAPERENode
SAPERENode
SAPERENode
SAPERENode
SAPERENode
SAPERENode
SAPERENode
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 12 / 46
Alchemist SAPERE background
Simulation of a SAPERE environment
The role of simulation
Emergence cannot be fully designed
It’s crazy to deploy a whole ecosystem without any test
Fist class abstractions
Dynamic environment
Different, mobile, communicating nodes
Programmability through a set of chemical-like laws
Continuous Time Markov Chain (CTMC) model
Autonomous agents
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 13 / 46
Alchemist SAPERE background
Two approaches
Classic ABM modelling
High flexibility
Topology as first-class abstraction
Dynamics explicitly modelled
No native support for CTMC model
Chemical inspired modelling
Natively CTMC
Very fast and reliable algorithms exist in literature
Very limited topology: multicompartment at best
Only classic reactions can change the world status: limited flexibility
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 14 / 46
Alchemist Computational model
Alchemist simulation approach
Base idea
Start from the existing work with stochastic chemical systemssimulation
Extend it as needed to reach desired flexibility
Goals
Full support for Continuous Time Markov Chains (CTMC)
Support for differently distributed events (e.g. Triggers)
Rich environments, with obstacles, mobile nodes, etc.
More expressive reactions
Fast and flexible SSA engine
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 15 / 46
Alchemist Computational model
Enriching the environment description
Environment
Node
Reactions
Molecules
Alchemist world
The Environment contains and links together Nodes
Each Node is programmed with a set of Reactions
Nodes contain Molecules
Each Molecule in a node is described with a Concentration
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 16 / 46
Alchemist Computational model
Extending the concept of reaction
From a set of reactants that combine themselves in a set of products to:
Number of
neighbors<3
Node
contains
something
Any other
condition
about this
environment
Rate equation: how conditions
influence the execution speed
Conditions Probability distribution Actions
Any other
action
on this
environment
Move a node
towards...
Change
concentration
of something
Reaction
In Alchemist, every event is an occurrence of a Reaction
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 17 / 46
Alchemist Engine and architecture
SSA Algorithms
Several SSA exist, they follow the same base schema [Gillespie, 1977]:
1 Select next reaction using markovian rates
2 Execute it
3 Update the rates which may have changed
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 18 / 46
Alchemist Engine and architecture
Do the math: reaction speed
Consider a chemical reaction in the form:
A + Bµ−→ C
The probability that this reaction will trigger depends on:
1 Number of molecules A and B: the higher, the higher the probabilitythose molecule will encounter and react
2 µ, a speed coefficient for the reaction
We define the propensity of a reaction as its speed in a precise instant oftime as
aµ = µ[A][B]
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 19 / 46
Alchemist Engine and architecture
Do the math: next reaction choice
If we assume every reaction is a Poisson process, the probability for it tobe the next one is:
P(next = µ) =
∫ ∞0
P(µ, τ)dτ =
∫ ∞0
aµe−τ
∑j ajdτ =
aµ∑j aj
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 20 / 46
Alchemist Engine and architecture
Do the math: next reaction time
We can also compute the next time of occurrence:
P(τ)dτ =∑j
P(µ, τ)dτ =
∑j
aj
e−τ∑
j ajdτ
∑j
aj = λ −→ λe−λx
F (x ≤ t) =
∫ t
−∞λe−λxdx =
[−e−λt
]t−∞
= e−λt
Now, given a uniformly distributed random r , it’s possible to compute it’sequivalent for the desired distribution:
e−λt = r ⇒ t =− ln (r)
λ
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 21 / 46
Alchemist Engine and architecture
Existing algorithms
Direct method
1 Compute propensity for each reaction
2 Select next reaction probabilistically
3 Execute it
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 22 / 46
Alchemist Engine and architecture
Existing algorithms
A+B→C
B+C→D E+G→A
D+E→E+F F→D+G
Direct method + Dependency graph
1 Compute the dependencies among reactions
2 Compute propensity for each reaction
3 Select next reaction probabilistically
4 Execute it
5 Update propensities only for potentially involved reactions
6 Goto 3Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 23 / 46
Alchemist Engine and architecture
Existing algorithms
Next Reaction
1 Compute a putative execution time for each reaction
2 Store reactions in a binary heap
3 Pick the next reaction
4 Execute it
5 Compute putative times only for potentially involved reactions
6 Goto 3
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 24 / 46
Alchemist Engine and architecture
Existing algorithms
[Slepoy et al., 2008]
1 Compute propensities
2 Split the reactions in groups by their propensity
3 Throw randoms until a reaction in a group is selected
4 Execute it
5 Update propensities only for potentially involved reactions
6 Goto 3
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 25 / 46
Alchemist Engine and architecture
More flexibility!
What they miss is what we added
Support for instantaneus events (∞ Markovian rate)
Reactions can be added and removed during the simulation
Support for non-exponential time distributed events (e.g. triggers)
Dependencies among reactions are evaluated considering their“context”, speeding up the update phase
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 26 / 46
Alchemist Engine and architecture
Smart data structures ⇒ bleeding edge performances
2.04 4
3.71
7.32 1
5.51 0
2
8.91 0
4.20 0
9.10 0
10.10 0
inf0 0
A+B→C
B+C→D E+G→A
D+E→E+F F→D+G
Next Reaction efficient structures made dynamic
Dynamic Indexed Priority Queue
Allow to access the next reaction to execute in O(1) timeWorst case update in log2 (N) (average case a lot better)Extended to ensure balancing with insertion and removal
Dynamic Dependency Graph
Allows to smartly update only a subset of all the reactionExtended with the concept of input and output context
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 27 / 46
Alchemist Engine and architecture
Modular structure
Environment
User Interface
Alchemist language
Application-specific Alchemist Bytecode Compiler
Environment description in application-specific language
Incarnation-specific language
Reporting System
Interactive UI
Reaction Manager
Dependency Graph
Core Engine
Simulation Flow Language Parser
Environment Instantiator
XML Bytecode
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 28 / 46
Alchemist Engine and architecture
Some Features in short
Parallel executor
Approximate Stochastic Model Checker
Parallelised engine
Alchemist2Blender
PVeStA integration
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 29 / 46
Alchemist Case studies
Crowd evacuation
EXIT 1
EXIT 2
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 30 / 46
Alchemist Case studies
Crowd steering
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 31 / 46
Alchemist Case studies
Adaptive Displays
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 32 / 46
Alchemist Case studies
Simple Morphogenesis proof of concept
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 33 / 46
Alchemist Case studies
Morphogenesis of a Drosophila Melanogaster
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 34 / 46
Alchemist Case studies
Anticipative adaptation
0 5
10 15
20 25
30 35
40 45 0
5
10
15
20
25
30
35
40
45
0
10
20
30
40
50
60
'datald.log' using 2:3:4
0
10
20
30
40
50
60
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 35 / 46
Alchemist Case studies
Realistic pedestrians
see the video...
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 36 / 46
Alchemist Performance
Testbed
We realised the same crowd-steering application for both Alchemistand RePast, in order to evaluate the performance gap (if any)
Source code for RePast and standalone Alchemist applicationavailable athttp://www.apice.unibo.it/xwiki/bin/view/Alchemist/JOS
We choose a simplified use case which allows simulation in RePastwithout any change in its engine
This actually cripples out the most important Alchemist optimizationfor complex environment (the dependency graph)
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 37 / 46
Alchemist Performance
Results
0
50
100
150
200
250
300
350
400
450
500
50 100 150 200 250 300 350 400 450 500
Exe
cuti
on t
ime [
s]
Number of agents
Performance comparison with Repast
Repast Alchemist
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 38 / 46
Development
Development
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 39 / 46
Development Development model
Distributed development
Alchemist is a training ground for some good team development practices
Linux kernel-like development model
Java
XText
Mercurial DCVS
Bitbucket web-based code hosting
Maven
JUnit
Source is released under GPL
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 40 / 46
Development Contributors
People involved
Michele Bombardi (Done)Realistic pedestrians
Chiara Casalboni (Done)Realistic pedestrians
Francesca Cioffi (Done)Further experiments with Alchemist-SAPERE
Davide Ensini (Done)Approximate Stochastic Model Checker improvement
Enrico Galassi (Done)Alchemist-SAPERE high level language
Enrico Gualandi (Ongoing)Alchemist-SAPERE languages review
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 41 / 46
Development Contributors
People involved
Luca Mella (Done)Tools for social network analysis
Alessandro Montalti (Done)MS-BioNet to AlchemistXML translator
Luca Nenni (Done)Alchemist2Blender
Enrico Polverelli (Done)Gradient based patterns
Michele Pratiffi (Done)Image importer for Alchemist
Giacomo Pronti (Done)SAPERE incarnation main author
Luca Ricci (Ongoing)Map importer for Alchemist
Andrea Vandin (IMT Lucca)PVeStA integration
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 42 / 46
Development Future activities
There is a lot of work to do!
MS-Bionet compatibility layer
AlcheGRID
Alchemist2Blender improvement
OpenStreet Map importer and Google Map importer
Blender Integration improvement
Gnuplot integration
CellML / SBML to AlchemistXML
Chemistry Incarnation review
Alchemist for Bio DSL
Realistic biological gradients
GPX tracks loader
RDF to Alchemist-SAPERE translator
These are examples, if you have something in mind, be proactive!Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 43 / 46
Development Future activities
Bibliography I
Banks, J. (1999).Introduction to simulation.In Farrington, P., Nembhard, H. B., Sturrock, D. T., and Evans,G. W., editors, Proceedings of the 1999 Winter SimulationConference, pages 7–13.
Gillespie, D. T. (1977).Exact stochastic simulation of coupled chemical reactions.Journal of Physical Chemistry, 81(25):2340–2361.
Klugl, F. and Norling, E. (2006).Agent-based simulation: Social science simulation and beyond.Technical report, The Eighth European Agent Systems SummerSchool (EASSS 2006), Annecy.
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 44 / 46
Development Future activities
Bibliography II
Parisi, D. (2001).Simulazioni - La realta rifatta al computer.Societa editrice il Mulino.
Slepoy, A., Thompson, A. P., and Plimpton, S. J. (2008).A constant-time kinetic Monte Carlo algorithm for simulation of largebiochemical reaction networks.The Journal of Chemical Physics, 128(20):205101+.
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 45 / 46
SAPERE Self-aware Pervasive Service Ecosystems
From Engineer to Alchemist, There and Back Again: AnAlchemist Tale
Danilo Pianini – [email protected]
Alma Mater Studiorum—Universita di Bologna
Cesena, May 24th, 2013
Danilo Pianini (UniBo) An Alchemist Tale Cesena, May 24th, 2013 46 / 46