Using TClass we have created a learner that can produce low error rates on classification when voted. On the Nintendo data, we are able to produce accuracy that is comparable to hand-selected set of features. On the Flock data, we attain an error of just 2 per cent.
When not voted, it can be used to create comprehensible descriptions that are easily understood in this context. The Nintendo data definitions are quite difficult to comprehend, mostly because of the limitations of the sensors. The Flock data produces definitions that can be compared (favourably) against definitions in the Auslan dictionary - perhaps not quite as comprehensible, but still, it is easy to see what the classifier is using.