A.1 Introduction

As a test and comparison to see how well algorithms coped with handshape recognition, we took a data set generously donated by Peter Vamplew, and applied the learning algorithms discussed here.

The tests were performed on the basic handshapes used in Auslan. Peter collected the information using a CyberGlove using only the finger information (i.e. not the information about wrist flexure). The features included were:

• Thumb's flexion towards the palm.

• Metacarpophalangeal (``first joint'') flexion for each of the five fingers.

• Proximal interphalangeal (``second joint'') flexion for each finger.

• Inter-finger abductions (i.e. thumb-index, index-middle, middle-ring, ring-little).

• Rotation of the little finger towards the palm.

In total, there were 16 features. Peter collected a total of ten data sets from ten different people, each with five samples of the variant handshapes. The data sets were divided into two sets, the unseen test cases (3) and the seen cases (7). Of the seen cases, 4 of the samples were used for training and the fifth for testing. Thus there were two test sets - seen candidates, and unseen candidates.