Inductive Logic Programming
Introduction to ILP
Inductive Logic Programming (ILP) is a research area formed at the intersection
of Machine Learning and Logic Programming. ILP systems develop predicate
descriptions from examples and background knowledge. The examples, background
knowledge and final descriptions are all described as logic programs. A
unifying theory of Inductive Logic Programming is being built up around
lattice-based concepts such as refinement, least general generalisation,
inverse resolution and most specific corrections. In addition to a well
established tradition of learning-in-the-limit results, some results within
Valiant's PAC-learning framework have been demonstrated for ILP systems.
U-learnabilty, a new model of learnability, has also been developed.
Presently successful applications areas for ILP systems include the
learning of structure-activity rules for drug design, finite-element mesh
analysis design rules, primary-secondary prediction of protein structure
and fault diagnosis rules for satellites.
Introduction to the Theory of ILP
of ILP is based on proof theory and model theory for the first order predicate
calculus. Inductive hypothesis formation is characterised by techniques
resolution , relative
least general generalisations , inverse
implication , and inverse
The removal of redundancy, and use of search procedures also play an
important role in the theory.
Computational Learning Theory (COLT) is used to analyse learning results
for ILP systems. An extension to Valiant's PAC learning framework, U-learnability,
has been suggested.
ILP systems have been applied to various problem domains. Many applications
benefit form the relational descriptions generated by the ILP systems.
The ability of ILP systems to accomodate background knowledge is also fundamental.
Some relationships learned in particular applications have been considered
as discoveries within those domains.
Application areas include: learning
drug structure-activity rules (drugs
for Alzheimer's disease, for
inhibition of E. Coli Dihydrofolate Reductase, in
suramin analogues); learning
rules for predicting mutagenesis; learning
rules for predicting protein secondary structure; learning
rules from chess databases; learning
rules for finite element mesh design; learning
diagnostic rules for qualitative models of satellite power supplies;
and learning qualitative
models of the U-tube system.
The applications are described in more detail on a separate
Implementations of ILP
This section introduces some implementaions of ILP.
The Duce system uses six inductive inference rules, of which, four are
concerened with definite clause propositional logic.
Cigol inverts resolution of clauses in first order predicate logic.
Golem is based on the construction of relative least general generalisations
Progol implements the theoretical framework of inverting entailment (IE).