Machine Learning for Emergent Middleware

Machine Learning for Emergent

Middleware

Amel Bennaceur, Issarny, Sykes (Inria)

Johansson and Moschitti (DISI)

Isberner, Steffen, Howar (TU Dortmund)

JIMSE 2012, Montpellier, France

The Interoperability Challenge

Pervasive networked systems dynamically discover each

other, combine functionality and interact to achieve goals

Highly heterogeneous protocols and data format

The Interoperability Challenge

Pervasive networked systems dynamically discover each

other, combine functionality and interact to achieve goals

Highly heterogeneous protocols and data format

Weather Station

Emergent Middleware to Support

Interoperability

Emergent

MiddlewareNetworked

System 2

Login

GetTemperature

GetHumidity

Logout

Networked

System 2

Login

GetWeather

Logout

Emergent

MiddlewareNetworked

System 2

Login

GetTemperature

GetHumidity

Logout

Networked

System 2

Login

GetWeather

Logout

Provides weather

Behaviour

Capability

Login

GetTemperature

GetHumidity

Logout

Requires weather

Behaviour

Capability

GetWeather

Logout

Login

wind

Atmosphere

Thing

Nothing

Ontology

weatherInfo

HumidityTemperature

Discovery

SystemLevel

ModelLevel

Learnin

gSynthesis Deployment

GetHumidityLogout

Login

GetWeather

Logout

Login

GetTemperature4

Mediator

Emergent

MiddlewareNetworked

System 2

Login

GetTemperature

GetHumidity

Logout

Networked

System 2

Login

GetWeather

Logout

Networked

System 1

Login

GetTemperature

GetHumidity

Logout

Learning the NS Model

Input

• The interface signature of the networked

system (NS)

Outputs

• The capability of the system: what the NS does

• The behaviour of the system:

how the NS achieves its capability

Algorithms

• Statistical learning for inferring NS capability

• Automata learning for extracting NS behaviour

Networked

System 1

Login

GetTemperature

GetHumidity

Logout

Provides weather

Behaviour

Capability

Login

GetTemperature

GetHumidity

Logout

??

Statistical Learning

for Inferring NS Capability

Classify systems into categories according

to their WSDL interface description

1. Apply standard techniques from text

categorisation

• Support Vector Machines

2. Extract the distinguishing features

• Use the bag-of-words representation: a

histogram of the words occurring in the

document to categorise

Statistical Learning

for Inferring NS Capability (example)

3. Represent a WSDL interface description as a

feature vector

• WSDL contains various types of “text”: comments, method

and parameter names

<wsdl:message name="GetWeatherByZipCodeSoapIn">

<wsdl:part name="parameters"

element="tns:GetWeatherByZipCode" />

</wsdl:message>

<wsdl:message name="GetWeatherByZipCodeSoapOut">

<wsdl:part name="parameters"

element="tns:GetWeatherByZipCodeResponse" />

</wsdl:message>

Extracted bag-of-words feature vector: get:4, weather:4, by:4, zip:4, code:4, soap:2, in:1, out:1, response:1

Automata Learning

for Extracting NS Behaviour

Incrementally construct a deterministic finite

automaton that matches the behaviour of the NS

on the basis of test-based interaction with the

system.

• Based on L* algorithm

• Start with the most general behaviour that allows any

sequence of the operations of the interface to be

executed

• Test and refine when an interaction errors, aka a

counterexample, is discovered

Automata Learning

for Extracting NS Behaviour (example)

GetWeatherLogin

Logout

t0

Networked

System 2

Login

GetWeather

Logout

Login Logout

GetWeather

t1

t2

Login

Logout

GetWeather

Counterexample 1

GetWeather Login

GetWeatherLogin Logout

Counterexample 2

The Benefits of Capability

Learned capabilities speed up matching

• Provides an efficient coarse-grained compatibility check, before

considering expensive behavioural checks

• Affordances must be equal or ontologically related (subclass)

The Cost of Behavioural Learning

Behavioural learning is still expensive

The Downside

Incorrect NS descriptions

• Prevent good connections from being made • matching or goal satisfaction fails

• Cause bad connections to be made • matching succeeds when it should not

• Cause good connections to fail through incorrect

synthesis

Feedback loop captures failures

Open Issue

Continuous learning to deal with evolution and incremental

synthesis

13

Monitoring

Evolution of the

NS Model

Discovery

Learning

Synthesis

New NSs

(re)Synthesis

Thank you

Further Information

Home page: www-rocq.inria.fr/~bennaceu

CONNECT: connect-forever.eu

The Role of Ontologies in Emergent Middleware:

Supporting Interoperability in Complex Distributed

Systems, In Proc. Middleware 2011

Middleware-layer Connector Synthesis: Beyond State of

the Art in Middleware Interoperability, In SFM 2011

Towards an architecture for runtime interoperability, In

Proc. ISoLA 2010

15