Machine Learning for Emergent Middleware - Trustworthy Eternal Systems via Evolving, Software Data and Knowledge - page 24

on the presence of underscores or CamelCase; all tokens are then normalised to
lowercase.
Once the feature representation is available, we use it to learn several classi-
fiers, each of them specialised to recognise if the WSDL expresses some target se-
mantic properties. The latter can also be concepts of an ontology. Consequently,
our algorithm may be used to learn classifiers that automatically assign ontol-
ogy concepts to actions defined in NS interfaces. Of course, the additional use of
domain (but at the same time general) ontologies facilitates the learning process
by providing effective features for the interface representation. In other words,
WSDL, domain ontologies and any other information contribute to defining the
vector representation used for training the concept classifiers.
5 Automata Learning for Inferring NS Behavioural
Semantics
Automata learning can be considered as a key technology for dealing with black
box systems, i.e., systems that can be observed, but for which no or little knowl-
edge about the internal structure or even their intent is available. Active Learn-
ing ( a.k.a regular extrapolation) attempts to construct a deterministic finite
automaton that matches the behaviour of a given target system on the ba-
sis of test-based interaction with the system. The popular L ∗ algorithm infers
Deterministic Finite Automata (DFAs) by means of membership queries that
test whether certain strings (potential runs) are contained in the target sys-
tem's language (its set of runs), and equivalence queries that compare intermedi-
ately constructed hypothesis automata for language equivalence with the target
system.
In its basic form, L ∗ starts with a hypothesis automaton that treats all se-
quences of considered input actions alike, i.e., it has one single state, and refines
this automaton on the basis of query results, iterating two main steps: (1) refining
intermediate hypothesis automata using membership queries until a certain level
of “consistency” is achieved ( test-based modelling ), and (2) testing hypothesis
automata for equivalence with the target system via equivalence queries ( model-
based testing ). This procedure successively produces state-minimal deterministic
(hypothesis) automata consistent with all the encountered query results [3]. This
basic pattern has been extended beyond the domain of learning DFAs to classes
of automata better suited for modelling reactive systems in practice. On the basis
of active learning algorithms for Mealy machines, inference algorithms for I/O-
automata [1], timed automata [7], Petri Nets [6], and Register Automata [10],
i.e., restricted flow graphs, have been developed.
While usually models produced by active learning are used in model-based
verification or some other domain that requires complete models of the system
under test (e.g., to prove absence of faults), here the inferred models serve as
a basis for the interaction with the system for Emergent Middleware synthesis.
This special focus poses unique requirements on the inferred models (discussed
in detail in [9]), which become apparent in the following prototypical example.