The simplest test possible is to feed the handshape data directly to the learning algorithms with no pre-processing.
Table A.1: Results of learning algorithms on Vamplew's handshape data
As can be seen, the IBLs outperform the C4.5 learners, but IBL2 and 3 do not seem to do as well as we would expect (previously they were only a few per cent behind IBL1). However, the results on seen users are quite good (95 per cent correct). It is easy to see how if we could get the right handshape 95 per cent of the time, we could use this information to further increase the accuracy of the sign recognition, by adding an additional feature specifying handshape.