Artificial Intelligence Research Group
Some Past Projects
Under Development
| 7. Multi-Strategy Constructive Induction |
| Investigator(s): Dr Kamal Nazar and Professor Max Bramer |
| Constructive induction attempts to construct new features, attributes and terms to alter the representation of a learning task. In practice deductive rules are used in conjunction with background knowledge when no specific overlap of features between example and concept definitions occurs. This transformation of representation is important because many types of learning problem are intractable, or practically intractable without a change in representation. Although it is possible to transform representation using recombination, and implementing this is relatively simple, recombination is an unacceptably complex task due to the large number, varied type and high arity of the constructive operators.
This research attempts to overcome the representational problems pathological to the quality of the concepts learned. If poor results are obtained initially, the inductive systems under development in this project perform automatic, problem orientated transformation of the representation space under consideration.
A loosely coupled approach has been adopted, initially coupling the Cupid framework as the inductive algorithm, and Multi-strategy attribute construction, but ensuring that the system is constructed so that other inductive algorithms can be used where appropriate or desirable. Features are constructed to improve the accuracy and decrease the complexity of the hypothesis generated during induction, and new attributes are constructed using the analysis of the induced hypothesis. Information theoretic and statistical measures are used in this analysis. Hypothesis ordering is used to constrain the complexity of the search and reduce the cost of evaluation.
|
|
Next
|
