Next: References
Up: A GBML Approach to
Previous: Result
We conclude the following:
- GA technique has been successful in discovering DFA.
- Hence it is
possible
to discover a regular grammar for a given language (if it exists).
- The `fitness function' serves multi-purpose optimizations well.
-
In this work the fitness function is used for
- seeking an accepting DFA,
- forcing minimization of the number of transitions in the DFA,
- forcing removal of looping transitions.
The result was observed to be very satisfactory.
- Providing only positive examples leads to a discovery of a
`more
generous' grammar. It is observed that although a
success of 100% acceptance of the testbed is achieved the
grammar `discovered' will generate additional strings which
would not be accepted by the grammar that is used in the
generation of the testbed. This is by no means an indications
of a weakness of the method. Being exposed to the
same example strings (from the testbed) a human intelligence
would presumably infer the some similar `generous' grammar.
The key point here is the need for negative examples. Examples
of strings that shall be rejected by the `discovered' DFA.
Although this would enhence the learning, it is considered as a
cheap method which beats the purpose. If such a clear cut knowledge of
what is correct, what is wrong is to hand, presumably the rules are
already known. Furthermore, positive examples are practically of infinite
amount if all the written corpus knowledge is considered. No such
knowledge is present in a corpus form regarding the negative
cases. Gathering it would need consulting a native speaker.
Meltem TURHAN
Tue Oct 29 22:25:58 EET 1996