The main long term technical goal of the ILP project is to upgrade the techniques of the classical empirical learning paradigm to a logic programming framework. In this way ILP aims to overcome the two main limitations of classical empirical or similarity based learning algorithms, such as the TDIDT-family: the use of a limited knowledge representation formalism (essentially a propositional logic), and the inability to use substantial background knowledge in the learning process.
The inductive logic programming (ILP) project has contributed theoretical foundations, both logical and complexity-theoretical, which were needed to provide further insight into the major challenges facing ILP, as well as provided fundamental techniques. The objectives of the ILP project were met and exceeded in each of the major topics addressed: theory revision, predicate invention, imperfect data handling, declarative bias, and ILP theory (covering both learnability issues and representational issues).
Notably, not only were theoretical results obtained, but systems based on these results were implemented and tested on a variety of problems,including some important real-world applications.
Some of the systems developed are freely available in the public domain for academic purposes.
The technology developed has been applied to important biological, ecological, engineering and business problems. For example, the Oxford results in several biological domains were published in respected biological and artificial intelligence journals, they were obtained using the general purpose ILP system Golem, and they were judged understandable by human biologists, a result which has seldom been achieved within artificial intelligence.
Besides data mining the other key application area of ILP technology lies in a software engineering context; some of the project results include automatic test-case generation and induction of loop invariants.
Theoretical results obtained address: the (non)-pac-learnability of certain classes of logic programs;
the study of an alternative semantics or problem specification for inductive logic programming (ILP) based on Helft's framework.
Results in theory revision included:
the development of multiple predicate learners in an incremental and empirical setting;
the study of minimal revisions to theories;
a technique to derive full clausal theories from deductive databases.
A better understanding of predicate invention was obtained:
by Muggleton's formal framework (and lattice) for predicate invention;
by introducing new techniques for predicate invention;
by comparative studies of predicate invention.
Results on handling imperfect data included:
adaptation of some mechanisms from attribute value learning to ILP;
development of stochastic ILP algorithms;
new results on information compression and Kolmogorov complexity.
Results on declarative bias included:
abstract frameworks for formulating bias and shifting the bias;
comparisons between existing frameworks for bias;
application of bias to programming assistants.
APPROACH AND METHODS
The project focuses on the following research topics:
- Theory of ILP: the theoretical implications of the use of logic programming for inductive learners.
This involves the study of:
. the properties of generalisation and specialisation operators such as inverse resolution
. the complexity and convergence aspects of particular inductive algorithms (this is concerned with learnability theory)
. logical frameworks for induction
. the development of a framework and methodology for empirical evaluation of ILP-learners.
- Theory Revision: the issues involved in learning multiple concepts in a first-order logic framework. Learning multiple concepts is a form of theory revision, where several related predicates or concepts may be modified or revised.
- Imperfect data: to upgrade and adapt existing noise-handling mechanisms form attribute value learning algorithms.
- Predicate Invention: the investigation of methods to invent new predicates. These methods aim at extending the vocabulary of the learner whenever the available vocabulary is unsatisfactory or insufficient and by doing so they extend the range of learnable concepts.
- Declarative Bias: the exploration of methods and formalisms to explicitly and declaratively represent the bias of inductive logic learners.
The expected outcome of the project is a sound basis for the development of systems that are able to induce logic programs from examples in real-life applications that involve substantial amounts of background knowledge.
53754 Sankt Augustin
106 91 Stockholm
OX1 3QD Oxford