Machine Learning for Emergent Middleware - Trustworthy Eternal Systems via Evolving, Software Data and Knowledge - page 27

hand side an abstract parameterised symbol is denoted, while on the right hand
side the named variable storing the invocation result is specified. Figure 5 de-
picts the behavioural description of the weather station, which was learned in
31 seconds on a portable computer, using 258 MQs.
The model correctly reflects the steps necessary, e.g., to read sensor data:
createProperties , createSession , getWeatherStation , authenticate and
getSensor have to be invoked before getSensorData can be called successfully.
Additionally, the actual realisation of authentication, which cannot be deduced
from the interface specification alone, is revealed in the inferred model. When
simply looking at the parameter types, the action getSensor should be invocable
directly after the getWeatherStation primitive. However, in reality getSensor
is guarded by an authentication mechanism, meaning that authenticate has to
be successfully invoked beforehand. Also, from the model, it is easily deducible
that the authenticate action will indeed merely affect the provided station
data object (and not, e.g., the whole session): requesting a new station data
object will always necessitate another authentication step before getSensor can
be invoked again, as that action requires an authenticated station data object.
6Conluon
This paper has presented the central role of learning in supporting the concept
of Emergent Middleware, which revisits the middleware paradigm to sustain in-
teroperability in increasingly heterogeneous and dynamic complex distributed
systems. The production of Emergent Middleware raises numerous challenges,
among which dealing with the apriori minimal knowledge about networked
systems that is available to the generation process. Indeed, semantic knowledge
about the interaction protocols run by the Networked Systems is needed to be
able to reason and compose protocols in a way that enable NSs to collabo-
rate properly. While making such knowledge available is increasingly common
in Internet-worked environments (e.g., see effort in the Web service domain), it
remains absent from the vast majority of descriptions exposed for the Networked
Systems that are made available over the Internet. This paper has specifically
outlined how powerful learning techniques that are being developed by the scien-
tific community can be successfully applied to the Emergent Middleware context,
thereby enabling the automated learning of both functional and behavioural se-
mantics of NSs. In more detail, this paper has detailed how statistical and au-
tomata learning can be exploited to enable on-the-fly inference of functional and
behavioural semantics of NSs, respectively.
Our experiments so far show great promise with respect to the effectiveness
and eciency of machine learning techniques applied to realistic distributed
systems such as in the GMES case. Our short-term future work focuses on the fine
tuning of machine learning algorithms according to the specifics of the networked
systems as well as enhancing the learnt models with data representations and
non-functional properties, which can result in considerable gains in terms of
accuracy and performance. In the mid-term, we will work on the realisation of