After my thesis, I became more involved with the machine learning project. Problems maintaining the portability of the original Weka workbench led to the decision in late 1997 to implement a new machine learning workbench entirely in Java. Over the last two years I have been responsible for managing the development of the Java workbench, written primarily by myself and two others. I have implemented machine learning schemes from instance-based learners, naive-Bayesian learners, and decision tree learners, to recent meta-schemes such as AdaBoost and Logitboost that improve the predictions of a base learner. To facilitate experimental evaluation of machine learning schemes, I developed an experiment architecture that permits experiments to be distributed over multiple machines, with results submitted to a central database where they may be combined and re-used for similar future experiments. The architecture incorporates dependencies and scheme versioning.

As well as participating in the development of the workbench, I interacted with research groups in horticultural and agricultural research institutes. These projects ranged from investigating the use of machine learning for objective grading of mushrooms, to the development of software sensors for process control.