Machine Learning for Emergent Middleware - Trustworthy Eternal Systems via Evolving, Software Data and Knowledge - page 22

categories were not given a priori. Obviously, the results of unsupervised
learning cannot compete with those of supervised learning.
- Semi-supervised learning is a pragmatic compromise. It allows one to use
a combination of a small labelled example set T s = { ( x, y ) } together with

in order to improve on both the
plain supervised learner making use of T s only and the unsupervised learner
using all available examples.
- Active learning puts the supervisor in a feedback loop: whenever the (active)
learner detects a situation where the available test set is inconclusive, the
learner actively constructs complementing examples and asks the supervisor
for the corresponding labelling. This learning discipline allows a much more
targeted learning process, since the active learner can focus on the impor-
tant/dicult cases (see for example [5]). The more structured the intended
learning output is, the more successful active learning will be, as the required
structural constraints are a good guide for the active construction of exam-
ples [3]. It has been successfully used in practice for inferring computational
models via testing [11,10].

a larger unlabelled example set T u = {x}

Learning technology has applicability in many domains. The next sections con-
centrate on the learning-based techniques that we are developing to enable the
automated inference of semantic knowledge about Networked Systems, both
functional and behavioural. The former relies on statistical learning while the
latter is based on automata learning .

4 Statistical Learning for Inferring NS Functional
Semantics

As discussed in Section 2.2, the first step in deciding whether two NSs will be
able to interoperate consists in checking the compatibility of their affordances
based on their functional semantics (i.e., ontology concepts characterising the
purpose of the affordance). Then, in the successful cases, behavioural matching
is performed so as to synthesise required mediator. This process highlights the
central role of the functional matching of affordances in reducing the overall com-
putation by acting as a kind of filter for the subsequent behavioural matching.
Unfortunately, legacy applications do not normally provide affordance descrip-
tions. We must therefore rely upon an engineer to provide them manually, or
find some automated means to extract the probable affordance from the interface
description. Note that it is not strictly necessary to have an absolutely correct
affordance since falsely-identified matches will be caught in the subsequent de-
tailed checks.
Since the interface is typically described by textual documentation, e.g., XML
documents, we can capitalise on the long tradition of research in text categorisa-
tion . This studies approaches for automatically enriching text documents with
semantic information. The latter is typically expressed by topic categories: thus
text categorisation proposes methods to assign documents (in our case, interface