The results of the histogram size tests are somewhat surprising, and difficult to decipher. There seems to be a general downward pattern up to about 4, but it is not consistently decreasing by any means after 4 divisions. A reasonable ``middle ground'' appears to be around the 6-division histograms. While not the best, the behaviour is predictable up to 6 and is in all cases except one, no worse by greater than 2 per cent than the minimum. There is of course some noise in the data, and this choice of 6 is reasonably arbitrary.
Another pattern is arising. While with up to three divisions, C4.5 is able to keep up, IBL1 outperforms it significantly beyond this level, by margins of around ten per cent. What may be happening is that the IBLs are better than C4.5 at handling the high dimensionality of the feature space. Since C4.5 can only consider the value of one variable at each node in the decision tree, this forces it to segment the space in limited ways relative to the way instance-based learning can.