Zulu is an active learning competition. Participants are to build
algorithms that can learn deterministic finite automata (DFA) by making
the smallest number of membership queries to the server/oracle.
Motivations
When learning language models, techniques usually make use of huge corpora
that are unavailable in many less resourced languages (such as the Zulu
language). One possible way around this problem is to interrogate an
expert with a number of chosen queries, in an interactive mode, until a
satisfying language model is reached. In this case, an important indicator
of success is the amount of energy the expert has spent in order for
learning to be successful. A nice learning paradigm covering this
situation is that of Query Learning, introduced by Dana Angluin.
In the field of Grammatical Inference, Query Learning was thoroughly
investigated to learn deterministic finite automata (DFA). As negative
results, it was proved that DFA could not be learned from just a
polynomial number of membership queries nor from just a polynomial number
of strong equivalence queries. On the other hand, algorithm L* designed by
Angluin, was proved to learn DFA from a polynomial number of both
membership and equivalence queries. These results yield several
successfull applications in Robotics, Games and Agents Technologies,
Information Retrieval, Hardware and Software Verification.
However, what has not been hardly studied is how to optimise the learning
task by trying to minimize the number of queries while making queries for
which the Oracle's work and answers are simple. These are strong
motivations for stemming research in the direction of developing new
interactive learning strategies and algorithms, that is the aim of this
competition.
Participation
The competition itself will be held in the spring of 2010.
Participants are encouraged to present their
innovations either as full papers to the ICGI 2010 conference, or as
extended abstracts to the Zulu workshop that will be organised during
ICGI 2010.
To know more go to the Zulu website [here]