Comprehensibility

Once the above high pruning criteria were introduced, the definitions became much clearer, even in the high noise case.

Figure 6.21: A decision tree produced by TClass on TTest with no noise. It is a perfect answer; ignoring gamma altogether as a noisy channel and having the absolute minimum number of nodes.

Figure 6.22: Events used by the decision tree in Figure 6.21.

The definitions arrived at for the no-noise case shown in Figure 6.21 are exactly correct. It looks for the increasing alpha value distinctive of the C channel; as well as the fact that class A has a big decreasing timestep right in the middle.

But what happens as we add noise? Do we still get such high quality definitions? If we use the high pruning levels we used above, we can sustain good definitions while $g = 0.1$. The results are shown in Figure 6.23.

Figure 6.23: A decision list produced by TClass on TTest with 10 per cent noise.

The definitions here are still understandable. The first rule looks for the beta channel with no significant local minima ant at least one local maximum, indicating a B class. Similarly, it looks for the big local minimum of the C class in the gamma channel.

Figure 6.24: Events used by the decision tree in Figure 6.23.

Finally,we look at the effect when $g = 0.2$ as shown in Figure 6.25. For brevity, we have omitted the event index. Although it is hard to understand such a complicated tree, some understanding can still be gained. For instance if the beta channel has a local maximum in the middle, and a nearby local minimum, it is likely to be of class A.

Figure 6.25: A decision tree produced by TClass on TTest with 20 per cent noise.