Ian M. Whittley

The next release problem

Companies developing and maintaining complex software systems need to determine the features that should be added to their system as part of the next release. They will wish to select these features to ensure the demands of their client base are satisfied as much as possible while at the same time ensuring that they themselves have the resources to undertake the necessary development. This situation is modelled in this paper and the problem of selecting an optimal next release is shown to be NP-hard. The use of various modern heuristics to find a high quality but possibly suboptimal solution is described. Comparative studies of these heuristics are given for various test cases.

Learning Classifier System Ensembles With Rule-Sharing

This paper presents an investigation into exploiting the population-based nature of learning classifier systems (LCSs) for their use within highly parallel systems. In particular, the use of simple payoff and accuracy-based LCSs within the ensemble machine approach is examined. Results indicate that inclusion of a rule migration mechanism inspired by parallel genetic algorithms is an effective way to improve learning speed in comparison to equivalent single systems. Presentation of a mechanism which exploits the underlying niche-based generalization mechanism of accuracy-based systems is then shown to further improve their performance, particularly, as task complexity increases. This is not found to be the case for payoff-based systems. Finally, considerably better than linear speedup is demonstrated with the accuracy-based systems on a version of the well-known Boolean logic benchmark task used throughout.

The Attribute Based Hill Climber

In this paper we introduce the Attribute Based Hill Climber, a parameter-free algorithm that provides a concrete, stand-alone implementation of a little used technique from the Tabu Search literature known as “regional aspiration”. Results of applying the algorithm to two classical optimisation problems, the Travelling Salesman Problem and the Quadratic Assignment Problem, show it to be competitive with existing general purpose heuristics in these areas.

On the use of rule-sharing in learning classifier system ensembles

This paper presents an investigation into exploiting the population-based nature of learning classifier systems for their use within highly-parallel systems. In particular, the use of simple accuracy-based learning classifier systems within the ensemble machine approach is examined. Results indicate that inclusion of a rule migration mechanism inspired by parallel genetic algorithms is an effective way to improve learning speed

Super Computer Heterogeneous Classifier Meta-Ensembles

This article describes the entry of the Super Computer Data Mining (SCDM) Project to the 10th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD) 2006 Data Mining Competition. The SCDM project is developing data mining tools for parallel execution on Linux clusters. The code is freely available; please contact the first author for a copy. We combine several classifiers, some of them ensemble techniques, into a heterogeneous meta-ensemble, to produce a probability estimate for each test case. We then use a simple decision theoretic framework to form a classification. The meta-ensemble contains a Bayesian neural network, a learning classifier system (LCS), attribute selection based-ensemble algorithms (Filtered At-tribute Subspace based Bagging with Injected Randomness [FASBIR]), and more well-known classifiers such as logistic regression, Naive Bayes (NB), and C4.5.

Variance Stabilizing Regression Ensembles for Environmental Models

This paper describes linear regression models fitted for the 2006 predictive uncertainty in environmental modelling competition hosted at the WCCI 2006 conference. Entries into this competition are required to produce models of up to four non-linear regression problems. Rather than adopt a complex non-linear modelling technique, our approach is to fit linear models to transformed data, with adaptive methods used for setting parameters and estimating error. This paper describes several techniques popular with statisticians which are less well known in the computational intelligence community, then proposes new ways of using these statistics. We describe standard statistical transformation techniques, Yeo-Johnson and Box-Tidwell, and present stepwise algorithms for using these transformations on large data sets. These stepwise algorithms utilise the Anscombe procedure, runs tests on residuals, the Goldfeld-Quandt procedure and the Kolomogorov-Smirnoff test for normality. We combine these statistics with the transformation procedures to form a piecewise linear approach to environmental modelling.