5.3.10 Other approaches considered but untested

There were of course, several other approaches that were considered. But these were eliminated for a number of reasons:

• A hierarchical approach to recognition. In this situation, the idea would have been to recognise the handshape using C4.5 or IBL, and then use the handshape decided upon as input into a ``second layer'' learner, again either C4.5 or IBL. However, it is unlikely this would have have worked for a number of reasons. As you can see in appendix A, the results of handshape recognition using the PowerGlove are not good at all. On average, the error rate for handshape recognition with the PowerGlove was 70 per cent. This would introduce significant noise into subsequent layers, which would probably have resulted in an increased error rate.

• A ``spatiotemporal encoding'' approach. In this sort of approach, the movement is broken into a series of component movements along a limited set of axes (eg up, down, left, right, forward, backward). These are output as strings and then compared at a string level. This was not used because of complexity and its vulnerability to noise.

We now consider ways to assemble these attributes to form a more accurate system.