RIBL (GMD)

System:	RIBL
Version:
Further specification:	empirical ILP system
Pointers:	ftp://ftp.mlnet.org/ml-archive/GMD/papers/ML79.ps.gz
Code:	Prolog executable, running on Sun SparcStations
References:	(Emde and Wettschereck 1996)
Other comments:	Can be used within the MOBAL system

RIBL is a relational instance-based learning (IBL) algorithm combining the advantages of instance-based/nearest-neighbor methods with a powerful relational (ILP) representation. As most instance-based systems, RIBL is robust, tolerates noise well and can handle numerical data as easily as symbolic data. Its similarity metric contains the usual propositional similarity metrics as a special case, so on propositional input, RIBL achieves the same quality of results known from propositional IBL learners. In contrast to most ILP systems, it cannot generate explicit sets of clauses as learning results.

RIBL can be used within the graphical environment of MOBAL or simply as a stand-alone command-line program. The program is composed of four main modules:

• A case generation module constructs descriptions of cases (i.e., a conjunction of literals) that serve as examples for learning or as test cases for classification.
• A module that computes the similarity of pairs of cases.
• A module that computes the relevance of predicates and their arguments.
• A module that implements distance-weighted k-nearest neighbor learning.

The learning input of RIBL consists of

• facts (unit clauses) that state properties of objects and relations among objects in the domain,
• predicate declarations that define the arity of the predicates as well as the sorts and mode declarations for their arguments, and
• type declarations that classify the terms that appear as arguments of facts into values and names, and declare the representation of values.

While facts and predicate declarations (including sorts) are representational constructs of MOBAL, type declarations are supported by RIBL. Type definitions are used to differentiate between arguments that represent an object in a domain (e.g., a person or an error code) and attribute values (e.g., a number that specifies the age of a person or the price of a component). In addition, type declarations specify how attribute values are represented (i.e., by using integers, reals, sets of unordered or ordered atoms). Mode declarations specify which arguments of a predicate are input or output arguments. This information is used in the same manner as it is used in other ILP learners (e.g., FOIL or GOLEM).

References

1. Werner Emde and Dietrich Wettschereck. Relational Instance Based Learning. In Lorenza Saitta, editor, Machine Learning - Proceedings 13th International Conference on Machine Learning, pages 122 - 130. Morgan Kaufmann, 1996.

back to index