An inductive method of system building is one that produces a generalized computational system from a set of specific examples. Neural computing offers one general class of inductive method, and rule-induction algorithms provide another. From a set of examples, a rule-induction algorithm will construct a decision tree, which is a computational system that embodies general rules that encapsulate the classifications (predictions, decisions, diagnoses, etc.) provided in the set of examples. Just as with neural networks, a decision tree cannot always be constructed, and when it can it is unlikely to be 100% correct even when classifying the training examples.
Decision trees tend to train faster than neural networks, but they tend to be more dificult to train to low error because they are more sensitive than neural networks to the choice of input features. However, the final decision tree is structured as sequences of decisions at the cognitive level; therefore, these computational systems may admit an interpretation and an understanding of the computational details of the problem in terms of cognitive-level features.
Figure 3 presents part of a decision tree for predicting whether an aircraft will level off at the flight level it is approaching or not. Such trees have been automatically generated using several thousand examples of aircraft leveling and not leveling at various flight levels at London's Heathrow Airport. The examples consisted of radar tracks (sampled at 4-sec intervals) and the particular flight level that was being approached (e.g. 7000-ft level, which is FL7000). Various features are extracted from the radar tracks; for example, in Fig. 3 d45 is the vertical distance traveled by the aircraft between radar points at interval 4 and interval 5 before the flight level is reached.
The internal structure of this (simplified) decision tree is composed of branch choices, and the tree leaves are the predicted outcomes (i.e., level off or non-level off). The branch choice at the top of the illustrated tree tests feature d45: If the value of d45 for the sample to be predicted is less than or equal to 100 feet, then the left branch is taken; otherwise, the right branch is taken and non-level off is immediately predicted.
Was this article helpful?