Peter Vamplew reported that he got results significantly better than the above using neural nets. His results were on a fully-connected, back-propagating neural network with 16 input nodes, 40 hidden nodes and 30 output nodes. His results are shown below.
Table A.3: Results of neural nets on Vamplew's handshape data
with additional attributes.
Obviously, these results are pretty good, although they are closely comparable with the results from IBL1. The only drawback is the learning time is quite high and the complexity of the neural net likewise, with a total of 86 nodes, 40 with 16 inputs and 30 with 40 inputs.