| Inductive Logic Programing and Automatic Programming: Towards Three Approaches. Popelinsky L, Flener P., Stepankova O.: Proceedings of 4th Int. ILP'94 workshop, Bonn 1994,Germany
Object-oriented data modelling and rules: ILP meets databases. Proceedings of Knowledge Level Modelling Workshop, ECML'95 Heraklion, Crete
WiM : A Study On Top-Down ILP Program. Popelinsky L., Stepankova O.: In: Proceedings of AIT'95 Workshop, Brno, Czech Republic, ISBN 80-214-0673-9
|
| The data set contains the minimal example sets for learning basic list processing predicates (member/2, append/3, reverse/2, split/3, delete/3, sublist/2, etc.), for the set operation union/3 as well as for predicates that use Peano's arithmetics ( plus/3, lessOrEqual/2, listLength/2, extractNth/3) by the system WiM [PopFleSte94, Popelinsky 95b].
For each target predicate, the example set contains the worst possible examples such that WiM can learn the target predicate. For the current version of WiM "the worst possible examples" have to meat the following three requirements:
- an instance of each base clause has to appear in the example set (e.g. for the predicate member/2 it may be member(a,[a,b])).
- the positive examples are not lying on the same resolution chain (e.g. member(d,[c,d,e]) is enough together with member(a,[a,b])) and
- at most one negative example is used (this negative example is generated by WiM itself).
Experiments with WiM to rebuild the database schema were described in [Popelinsky 95a]. In deductive object-oriented databases both classes and attributes may be defined by rules. Example sets for learning subclasses (japaneseCar/1 as a subclass of car/1, isMother/1 as of person/1, electricalVehicle/1 as a subclass of car/1 and publicTransportVehicle/1), superclasses (factory/1, person/1) as well as example sets for learning classes of new objects (family/2) and for learning a new attribute (personManagedBy/2) are included.
|